diff --git "a/data/refined_40k/test_bench/data_inc.json" "b/data/refined_40k/test_bench/data_inc.json" new file mode 100644--- /dev/null +++ "b/data/refined_40k/test_bench/data_inc.json" @@ -0,0 +1,12610 @@ +[ + { + "id": "7001917", + "images": [ + "7001917_00.jpg", + "7001917_01.jpg", + "7001917_02.jpg", + "7001917_03.jpg", + "7001917_04.jpg", + "7001917_05.jpg", + "7001917_06.jpg", + "7001917_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"crowd surfing at a Pink Floid rock concert before the show \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001486", + "images": [ + "7001486_00.jpg", + "7001486_01.jpg", + "7001486_02.jpg", + "7001486_03.jpg", + "7001486_04.jpg", + "7001486_05.jpg", + "7001486_06.jpg", + "7001486_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a robot holding baby Jesus at the nativity \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "0001611", + "images": [ + "0001611_00.jpg", + "0001611_01.jpg", + "0001611_02.jpg", + "0001611_03.jpg", + "0001611_04.jpg", + "0001611_05.jpg", + "0001611_06.jpg", + "0001611_07.jpg", + "0001611_08.jpg", + "0001611_09.jpg", + "0001611_10.jpg", + "0001611_11.jpg", + "0001611_12.jpg", + "0001611_13.jpg", + "0001611_14.jpg", + "0001611_15.jpg", + "0001611_16.jpg", + "0001611_17.jpg", + "0001611_18.jpg", + "0001611_19.jpg", + "0001611_20.jpg", + "0001611_21.jpg", + "0001611_22.jpg", + "0001611_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Near the Elephant Gate village, they approach the haunted house at night. Rajiv feels anxious, but Bhavesh encourages him. As they reach the house, a mysterious sound in the air adds to the suspense. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000063", + "images": [ + "3000063_00.jpg", + "3000063_01.jpg", + "3000063_02.jpg", + "3000063_03.jpg", + "3000063_04.jpg", + "3000063_05.jpg", + "3000063_06.jpg", + "3000063_07.jpg", + "3000063_08.jpg", + "3000063_09.jpg", + "3000063_10.jpg", + "3000063_11.jpg", + "3000063_12.jpg", + "3000063_13.jpg", + "3000063_14.jpg", + "3000063_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Homer Simpson is running towards you down a hallway in the back rooms \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "4000615", + "images": [ + "4000615_00.jpg", + "4000615_01.jpg", + "4000615_02.jpg", + "4000615_03.jpg", + "4000615_04.jpg", + "4000615_05.jpg", + "4000615_06.jpg", + "4000615_07.jpg", + "4000615_08.jpg", + "4000615_09.jpg", + "4000615_10.jpg", + "4000615_11.jpg", + "4000615_12.jpg", + "4000615_13.jpg", + "4000615_14.jpg", + "4000615_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Scene of burning incense and sipping tea in a Chinese tea room \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000835", + "images": [ + "5000835_00.jpg", + "5000835_01.jpg", + "5000835_02.jpg", + "5000835_03.jpg", + "5000835_04.jpg", + "5000835_05.jpg", + "5000835_06.jpg", + "5000835_07.jpg", + "5000835_08.jpg", + "5000835_09.jpg", + "5000835_10.jpg", + "5000835_11.jpg", + "5000835_12.jpg", + "5000835_13.jpg", + "5000835_14.jpg", + "5000835_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"dead tiger in forest with blood rodent cartoon style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "3002527", + "images": [ + "3002527_00.jpg", + "3002527_01.jpg", + "3002527_02.jpg", + "3002527_03.jpg", + "3002527_04.jpg", + "3002527_05.jpg", + "3002527_06.jpg", + "3002527_07.jpg", + "3002527_08.jpg", + "3002527_09.jpg", + "3002527_10.jpg", + "3002527_11.jpg", + "3002527_12.jpg", + "3002527_13.jpg", + "3002527_14.jpg", + "3002527_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"lighting strikes clouds all over hyper realistic a storm going on and a city aerial view of the storm zoomed in from the street \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001328", + "images": [ + "6001328_00.jpg", + "6001328_01.jpg", + "6001328_02.jpg", + "6001328_03.jpg", + "6001328_04.jpg", + "6001328_05.jpg", + "6001328_06.jpg", + "6001328_07.jpg", + "6001328_08.jpg", + "6001328_09.jpg", + "6001328_10.jpg", + "6001328_11.jpg", + "6001328_12.jpg", + "6001328_13.jpg", + "6001328_14.jpg", + "6001328_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Green field filled with flowers that sway gently in the wind \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000385", + "images": [ + "3000385_00.jpg", + "3000385_01.jpg", + "3000385_02.jpg", + "3000385_03.jpg", + "3000385_04.jpg", + "3000385_05.jpg", + "3000385_06.jpg", + "3000385_07.jpg", + "3000385_08.jpg", + "3000385_09.jpg", + "3000385_10.jpg", + "3000385_11.jpg", + "3000385_12.jpg", + "3000385_13.jpg", + "3000385_14.jpg", + "3000385_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"overhead drone footage of ocean shoreline with atlantic puffins nesting \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "1001382", + "images": [ + "1001382_00.jpg", + "1001382_01.jpg", + "1001382_02.jpg", + "1001382_03.jpg", + "1001382_04.jpg", + "1001382_05.jpg", + "1001382_06.jpg", + "1001382_07.jpg", + "1001382_08.jpg", + "1001382_09.jpg", + "1001382_10.jpg", + "1001382_11.jpg", + "1001382_12.jpg", + "1001382_13.jpg", + "1001382_14.jpg", + "1001382_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"the immense size of a blue whale heart comparing it to the dimensions of a small car \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000792", + "images": [ + "7000792_00.jpg", + "7000792_01.jpg", + "7000792_02.jpg", + "7000792_03.jpg", + "7000792_04.jpg", + "7000792_05.jpg", + "7000792_06.jpg", + "7000792_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"good looking man standing up to his chest in water at the beach \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "7000492", + "images": [ + "7000492_00.jpg", + "7000492_01.jpg", + "7000492_02.jpg", + "7000492_03.jpg", + "7000492_04.jpg", + "7000492_05.jpg", + "7000492_06.jpg", + "7000492_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The Prince on a white horse \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "2001481", + "images": [ + "2001481_00.jpg", + "2001481_01.jpg", + "2001481_02.jpg", + "2001481_03.jpg", + "2001481_04.jpg", + "2001481_05.jpg", + "2001481_06.jpg", + "2001481_07.jpg", + "2001481_08.jpg", + "2001481_09.jpg", + "2001481_10.jpg", + "2001481_11.jpg", + "2001481_12.jpg", + "2001481_13.jpg", + "2001481_14.jpg", + "2001481_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The villagers asked what the fish say \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "0001284", + "images": [ + "0001284_00.jpg", + "0001284_01.jpg", + "0001284_02.jpg", + "0001284_03.jpg", + "0001284_04.jpg", + "0001284_05.jpg", + "0001284_06.jpg", + "0001284_07.jpg", + "0001284_08.jpg", + "0001284_09.jpg", + "0001284_10.jpg", + "0001284_11.jpg", + "0001284_12.jpg", + "0001284_13.jpg", + "0001284_14.jpg", + "0001284_15.jpg", + "0001284_16.jpg", + "0001284_17.jpg", + "0001284_18.jpg", + "0001284_19.jpg", + "0001284_20.jpg", + "0001284_21.jpg", + "0001284_22.jpg", + "0001284_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Iron Man lands in old Marvel style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "4001069", + "images": [ + "4001069_00.jpg", + "4001069_01.jpg", + "4001069_02.jpg", + "4001069_03.jpg", + "4001069_04.jpg", + "4001069_05.jpg", + "4001069_06.jpg", + "4001069_07.jpg", + "4001069_08.jpg", + "4001069_09.jpg", + "4001069_10.jpg", + "4001069_11.jpg", + "4001069_12.jpg", + "4001069_13.jpg", + "4001069_14.jpg", + "4001069_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a man pulling handpump from the ground \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000970", + "images": [ + "4000970_00.jpg", + "4000970_01.jpg", + "4000970_02.jpg", + "4000970_03.jpg", + "4000970_04.jpg", + "4000970_05.jpg", + "4000970_06.jpg", + "4000970_07.jpg", + "4000970_08.jpg", + "4000970_09.jpg", + "4000970_10.jpg", + "4000970_11.jpg", + "4000970_12.jpg", + "4000970_13.jpg", + "4000970_14.jpg", + "4000970_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"wedding dragon party in a rave with several dragons dancing in a huge castle illuminate with strobbing rgb lights, fast traveling camera, ar 16:9 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "7001302", + "images": [ + "7001302_00.jpg", + "7001302_01.jpg", + "7001302_02.jpg", + "7001302_03.jpg", + "7001302_04.jpg", + "7001302_05.jpg", + "7001302_06.jpg", + "7001302_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a young 16 year old kid in Napoli jersey \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000117", + "images": [ + "4000117_00.jpg", + "4000117_01.jpg", + "4000117_02.jpg", + "4000117_03.jpg", + "4000117_04.jpg", + "4000117_05.jpg", + "4000117_06.jpg", + "4000117_07.jpg", + "4000117_08.jpg", + "4000117_09.jpg", + "4000117_10.jpg", + "4000117_11.jpg", + "4000117_12.jpg", + "4000117_13.jpg", + "4000117_14.jpg", + "4000117_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A futuristic science fiction scene of a monk standing in front of a quantum computer, in the stlye of roger deakins, cinematic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001465", + "images": [ + "7001465_00.jpg", + "7001465_01.jpg", + "7001465_02.jpg", + "7001465_03.jpg", + "7001465_04.jpg", + "7001465_05.jpg", + "7001465_06.jpg", + "7001465_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"fish swiming fast for his life \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002051", + "images": [ + "2002051_00.jpg", + "2002051_01.jpg", + "2002051_02.jpg", + "2002051_03.jpg", + "2002051_04.jpg", + "2002051_05.jpg", + "2002051_06.jpg", + "2002051_07.jpg", + "2002051_08.jpg", + "2002051_09.jpg", + "2002051_10.jpg", + "2002051_11.jpg", + "2002051_12.jpg", + "2002051_13.jpg", + "2002051_14.jpg", + "2002051_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"prison submerged in handcuffs and lemon juice and lemons \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002332", + "images": [ + "2002332_00.jpg", + "2002332_01.jpg", + "2002332_02.jpg", + "2002332_03.jpg", + "2002332_04.jpg", + "2002332_05.jpg", + "2002332_06.jpg", + "2002332_07.jpg", + "2002332_08.jpg", + "2002332_09.jpg", + "2002332_10.jpg", + "2002332_11.jpg", + "2002332_12.jpg", + "2002332_13.jpg", + "2002332_14.jpg", + "2002332_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"beautiful brilliant colors sunset on the shore of the beautiful ocean,8k \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000464", + "images": [ + "3000464_00.jpg", + "3000464_01.jpg", + "3000464_02.jpg", + "3000464_03.jpg", + "3000464_04.jpg", + "3000464_05.jpg", + "3000464_06.jpg", + "3000464_07.jpg", + "3000464_08.jpg", + "3000464_09.jpg", + "3000464_10.jpg", + "3000464_11.jpg", + "3000464_12.jpg", + "3000464_13.jpg", + "3000464_14.jpg", + "3000464_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Sudden appearance of shadows , contouring the walls in a fortaleza de Santa Cruz de Anhatomirim terror scene \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000319", + "images": [ + "7000319_00.jpg", + "7000319_01.jpg", + "7000319_02.jpg", + "7000319_03.jpg", + "7000319_04.jpg", + "7000319_05.jpg", + "7000319_06.jpg", + "7000319_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"In the style of Luchino Visconti, grainy black and white: Slow pan across an array of unearthed artifacts carefully laid on a cloth, their significance narrated through stark lighting \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001211", + "images": [ + "7001211_00.jpg", + "7001211_01.jpg", + "7001211_02.jpg", + "7001211_03.jpg", + "7001211_04.jpg", + "7001211_05.jpg", + "7001211_06.jpg", + "7001211_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"High resolution, a female mechanic works on her 60s muscle car, all that can be see are her beautiful legs and lower half of her torso, the other half is under the car working and is not seen, She is wearing ultra High Heeled shoes and short shorts, bright outdoor lighting \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000066", + "images": [ + "4000066_00.jpg", + "4000066_01.jpg", + "4000066_02.jpg", + "4000066_03.jpg", + "4000066_04.jpg", + "4000066_05.jpg", + "4000066_06.jpg", + "4000066_07.jpg", + "4000066_08.jpg", + "4000066_09.jpg", + "4000066_10.jpg", + "4000066_11.jpg", + "4000066_12.jpg", + "4000066_13.jpg", + "4000066_14.jpg", + "4000066_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"1800s freemasons military zoom in film \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000039", + "images": [ + "6000039_00.jpg", + "6000039_01.jpg", + "6000039_02.jpg", + "6000039_03.jpg", + "6000039_04.jpg", + "6000039_05.jpg", + "6000039_06.jpg", + "6000039_07.jpg", + "6000039_08.jpg", + "6000039_09.jpg", + "6000039_10.jpg", + "6000039_11.jpg", + "6000039_12.jpg", + "6000039_13.jpg", + "6000039_14.jpg", + "6000039_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"After many trials, fish finally reached the location marked on the map. cartoon style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "0002076", + "images": [ + "0002076_00.jpg", + "0002076_01.jpg", + "0002076_02.jpg", + "0002076_03.jpg", + "0002076_04.jpg", + "0002076_05.jpg", + "0002076_06.jpg", + "0002076_07.jpg", + "0002076_08.jpg", + "0002076_09.jpg", + "0002076_10.jpg", + "0002076_11.jpg", + "0002076_12.jpg", + "0002076_13.jpg", + "0002076_14.jpg", + "0002076_15.jpg", + "0002076_16.jpg", + "0002076_17.jpg", + "0002076_18.jpg", + "0002076_19.jpg", + "0002076_20.jpg", + "0002076_21.jpg", + "0002076_22.jpg", + "0002076_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"View from a plant and in the sky you can see more planets and some green and pink nebulae, camera in zoom Message: Aldhair (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001479", + "images": [ + "6001479_00.jpg", + "6001479_01.jpg", + "6001479_02.jpg", + "6001479_03.jpg", + "6001479_04.jpg", + "6001479_05.jpg", + "6001479_06.jpg", + "6001479_07.jpg", + "6001479_08.jpg", + "6001479_09.jpg", + "6001479_10.jpg", + "6001479_11.jpg", + "6001479_12.jpg", + "6001479_13.jpg", + "6001479_14.jpg", + "6001479_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Sage Kavita holding an ancient scroll with faded prophecies. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "2001934", + "images": [ + "2001934_00.jpg", + "2001934_01.jpg", + "2001934_02.jpg", + "2001934_03.jpg", + "2001934_04.jpg", + "2001934_05.jpg", + "2001934_06.jpg", + "2001934_07.jpg", + "2001934_08.jpg", + "2001934_09.jpg", + "2001934_10.jpg", + "2001934_11.jpg", + "2001934_12.jpg", + "2001934_13.jpg", + "2001934_14.jpg", + "2001934_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A man and a friend. Eating chips and watching a football game, waving his arms \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001094", + "images": [ + "4001094_00.jpg", + "4001094_01.jpg", + "4001094_02.jpg", + "4001094_03.jpg", + "4001094_04.jpg", + "4001094_05.jpg", + "4001094_06.jpg", + "4001094_07.jpg", + "4001094_08.jpg", + "4001094_09.jpg", + "4001094_10.jpg", + "4001094_11.jpg", + "4001094_12.jpg", + "4001094_13.jpg", + "4001094_14.jpg", + "4001094_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"smooth movement of the character, increasing the perspective of the face \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002934", + "images": [ + "1002934_00.jpg", + "1002934_01.jpg", + "1002934_02.jpg", + "1002934_03.jpg", + "1002934_04.jpg", + "1002934_05.jpg", + "1002934_06.jpg", + "1002934_07.jpg", + "1002934_08.jpg", + "1002934_09.jpg", + "1002934_10.jpg", + "1002934_11.jpg", + "1002934_12.jpg", + "1002934_13.jpg", + "1002934_14.jpg", + "1002934_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Alice in wonderland faking down an endless rabbit hole \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "8000719", + "images": [ + "8000719_00.jpg", + "8000719_01.jpg", + "8000719_02.jpg", + "8000719_03.jpg", + "8000719_04.jpg", + "8000719_05.jpg", + "8000719_06.jpg", + "8000719_07.jpg", + "8000719_08.jpg", + "8000719_09.jpg", + "8000719_10.jpg", + "8000719_11.jpg", + "8000719_12.jpg", + "8000719_13.jpg", + "8000719_14.jpg", + "8000719_15.jpg", + "8000719_16.jpg", + "8000719_17.jpg", + "8000719_18.jpg", + "8000719_19.jpg", + "8000719_20.jpg", + "8000719_21.jpg", + "8000719_22.jpg", + "8000719_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Scene 2: Show the animals happily living their lives until the drought strikes. 8k full hd \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001506", + "images": [ + "4001506_00.jpg", + "4001506_01.jpg", + "4001506_02.jpg", + "4001506_03.jpg", + "4001506_04.jpg", + "4001506_05.jpg", + "4001506_06.jpg", + "4001506_07.jpg", + "4001506_08.jpg", + "4001506_09.jpg", + "4001506_10.jpg", + "4001506_11.jpg", + "4001506_12.jpg", + "4001506_13.jpg", + "4001506_14.jpg", + "4001506_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Caillou sat on the throne of the blood god \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000182", + "images": [ + "4000182_00.jpg", + "4000182_01.jpg", + "4000182_02.jpg", + "4000182_03.jpg", + "4000182_04.jpg", + "4000182_05.jpg", + "4000182_06.jpg", + "4000182_07.jpg", + "4000182_08.jpg", + "4000182_09.jpg", + "4000182_10.jpg", + "4000182_11.jpg", + "4000182_12.jpg", + "4000182_13.jpg", + "4000182_14.jpg", + "4000182_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"psychedelic mandlbrot encapsulated in a glowing sphere, cymatics, ferrofluid \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002807", + "images": [ + "2002807_00.jpg", + "2002807_01.jpg", + "2002807_02.jpg", + "2002807_03.jpg", + "2002807_04.jpg", + "2002807_05.jpg", + "2002807_06.jpg", + "2002807_07.jpg", + "2002807_08.jpg", + "2002807_09.jpg", + "2002807_10.jpg", + "2002807_11.jpg", + "2002807_12.jpg", + "2002807_13.jpg", + "2002807_14.jpg", + "2002807_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Madonna Ciccone Song in the street \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002651", + "images": [ + "3002651_00.jpg", + "3002651_01.jpg", + "3002651_02.jpg", + "3002651_03.jpg", + "3002651_04.jpg", + "3002651_05.jpg", + "3002651_06.jpg", + "3002651_07.jpg", + "3002651_08.jpg", + "3002651_09.jpg", + "3002651_10.jpg", + "3002651_11.jpg", + "3002651_12.jpg", + "3002651_13.jpg", + "3002651_14.jpg", + "3002651_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A man with a Crown talking to his son. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "0001073", + "images": [ + "0001073_00.jpg", + "0001073_01.jpg", + "0001073_02.jpg", + "0001073_03.jpg", + "0001073_04.jpg", + "0001073_05.jpg", + "0001073_06.jpg", + "0001073_07.jpg", + "0001073_08.jpg", + "0001073_09.jpg", + "0001073_10.jpg", + "0001073_11.jpg", + "0001073_12.jpg", + "0001073_13.jpg", + "0001073_14.jpg", + "0001073_15.jpg", + "0001073_16.jpg", + "0001073_17.jpg", + "0001073_18.jpg", + "0001073_19.jpg", + "0001073_20.jpg", + "0001073_21.jpg", + "0001073_22.jpg", + "0001073_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The space Fleet is storming the planet \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "2002231", + "images": [ + "2002231_00.jpg", + "2002231_01.jpg", + "2002231_02.jpg", + "2002231_03.jpg", + "2002231_04.jpg", + "2002231_05.jpg", + "2002231_06.jpg", + "2002231_07.jpg", + "2002231_08.jpg", + "2002231_09.jpg", + "2002231_10.jpg", + "2002231_11.jpg", + "2002231_12.jpg", + "2002231_13.jpg", + "2002231_14.jpg", + "2002231_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Shifting universe, distorted, dislocation, splitting powerful image of a humanoid mecha \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002204", + "images": [ + "3002204_00.jpg", + "3002204_01.jpg", + "3002204_02.jpg", + "3002204_03.jpg", + "3002204_04.jpg", + "3002204_05.jpg", + "3002204_06.jpg", + "3002204_07.jpg", + "3002204_08.jpg", + "3002204_09.jpg", + "3002204_10.jpg", + "3002204_11.jpg", + "3002204_12.jpg", + "3002204_13.jpg", + "3002204_14.jpg", + "3002204_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"character dnd concept, Snow leopard like tabaxi paladin, great magical sword \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "0002425", + "images": [ + "0002425_00.jpg", + "0002425_01.jpg", + "0002425_02.jpg", + "0002425_03.jpg", + "0002425_04.jpg", + "0002425_05.jpg", + "0002425_06.jpg", + "0002425_07.jpg", + "0002425_08.jpg", + "0002425_09.jpg", + "0002425_10.jpg", + "0002425_11.jpg", + "0002425_12.jpg", + "0002425_13.jpg", + "0002425_14.jpg", + "0002425_15.jpg", + "0002425_16.jpg", + "0002425_17.jpg", + "0002425_18.jpg", + "0002425_19.jpg", + "0002425_20.jpg", + "0002425_21.jpg", + "0002425_22.jpg", + "0002425_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"3d cartoon animation video of two chicks one chick is cooking and the other is sitting next to him \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002932", + "images": [ + "2002932_00.jpg", + "2002932_01.jpg", + "2002932_02.jpg", + "2002932_03.jpg", + "2002932_04.jpg", + "2002932_05.jpg", + "2002932_06.jpg", + "2002932_07.jpg", + "2002932_08.jpg", + "2002932_09.jpg", + "2002932_10.jpg", + "2002932_11.jpg", + "2002932_12.jpg", + "2002932_13.jpg", + "2002932_14.jpg", + "2002932_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"young boy well dressed wearing belt in empty room looking ahead with confidence \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4001509", + "images": [ + "4001509_00.jpg", + "4001509_01.jpg", + "4001509_02.jpg", + "4001509_03.jpg", + "4001509_04.jpg", + "4001509_05.jpg", + "4001509_06.jpg", + "4001509_07.jpg", + "4001509_08.jpg", + "4001509_09.jpg", + "4001509_10.jpg", + "4001509_11.jpg", + "4001509_12.jpg", + "4001509_13.jpg", + "4001509_14.jpg", + "4001509_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"25 seconds of a crow flying \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001184", + "images": [ + "6001184_00.jpg", + "6001184_01.jpg", + "6001184_02.jpg", + "6001184_03.jpg", + "6001184_04.jpg", + "6001184_05.jpg", + "6001184_06.jpg", + "6001184_07.jpg", + "6001184_08.jpg", + "6001184_09.jpg", + "6001184_10.jpg", + "6001184_11.jpg", + "6001184_12.jpg", + "6001184_13.jpg", + "6001184_14.jpg", + "6001184_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a train 1960 going in south italy , autumn and windy weather close to the sea \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000530", + "images": [ + "3000530_00.jpg", + "3000530_01.jpg", + "3000530_02.jpg", + "3000530_03.jpg", + "3000530_04.jpg", + "3000530_05.jpg", + "3000530_06.jpg", + "3000530_07.jpg", + "3000530_08.jpg", + "3000530_09.jpg", + "3000530_10.jpg", + "3000530_11.jpg", + "3000530_12.jpg", + "3000530_13.jpg", + "3000530_14.jpg", + "3000530_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Running girl, full body, moving camera, weaving in and out of the city. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "0002895", + "images": [ + "0002895_00.jpg", + "0002895_01.jpg", + "0002895_02.jpg", + "0002895_03.jpg", + "0002895_04.jpg", + "0002895_05.jpg", + "0002895_06.jpg", + "0002895_07.jpg", + "0002895_08.jpg", + "0002895_09.jpg", + "0002895_10.jpg", + "0002895_11.jpg", + "0002895_12.jpg", + "0002895_13.jpg", + "0002895_14.jpg", + "0002895_15.jpg", + "0002895_16.jpg", + "0002895_17.jpg", + "0002895_18.jpg", + "0002895_19.jpg", + "0002895_20.jpg", + "0002895_21.jpg", + "0002895_22.jpg", + "0002895_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The property of the convicted shall not be confiscated, fast motion \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000217", + "images": [ + "5000217_00.jpg", + "5000217_01.jpg", + "5000217_02.jpg", + "5000217_03.jpg", + "5000217_04.jpg", + "5000217_05.jpg", + "5000217_06.jpg", + "5000217_07.jpg", + "5000217_08.jpg", + "5000217_09.jpg", + "5000217_10.jpg", + "5000217_11.jpg", + "5000217_12.jpg", + "5000217_13.jpg", + "5000217_14.jpg", + "5000217_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A robot that throws a stack of paper from a desk \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "1000299", + "images": [ + "1000299_00.jpg", + "1000299_01.jpg", + "1000299_02.jpg", + "1000299_03.jpg", + "1000299_04.jpg", + "1000299_05.jpg", + "1000299_06.jpg", + "1000299_07.jpg", + "1000299_08.jpg", + "1000299_09.jpg", + "1000299_10.jpg", + "1000299_11.jpg", + "1000299_12.jpg", + "1000299_13.jpg", + "1000299_14.jpg", + "1000299_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A close up of a beautiful Queen crying on the death of Alexander The great. Alexander the Great lying on the bed surrounded by his courtiers \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001767", + "images": [ + "6001767_00.jpg", + "6001767_01.jpg", + "6001767_02.jpg", + "6001767_03.jpg", + "6001767_04.jpg", + "6001767_05.jpg", + "6001767_06.jpg", + "6001767_07.jpg", + "6001767_08.jpg", + "6001767_09.jpg", + "6001767_10.jpg", + "6001767_11.jpg", + "6001767_12.jpg", + "6001767_13.jpg", + "6001767_14.jpg", + "6001767_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Her ability to unite people under a common purpose and her deep empathy for the struggles of her subjects made her a beloved ruler. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "7001972", + "images": [ + "7001972_00.jpg", + "7001972_01.jpg", + "7001972_02.jpg", + "7001972_03.jpg", + "7001972_04.jpg", + "7001972_05.jpg", + "7001972_06.jpg", + "7001972_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"animate the flying saucers and make the beard flow with the wind \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "4001467", + "images": [ + "4001467_00.jpg", + "4001467_01.jpg", + "4001467_02.jpg", + "4001467_03.jpg", + "4001467_04.jpg", + "4001467_05.jpg", + "4001467_06.jpg", + "4001467_07.jpg", + "4001467_08.jpg", + "4001467_09.jpg", + "4001467_10.jpg", + "4001467_11.jpg", + "4001467_12.jpg", + "4001467_13.jpg", + "4001467_14.jpg", + "4001467_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"lilliputs playing in doll house, 3d , realistic , 4k \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002156", + "images": [ + "2002156_00.jpg", + "2002156_01.jpg", + "2002156_02.jpg", + "2002156_03.jpg", + "2002156_04.jpg", + "2002156_05.jpg", + "2002156_06.jpg", + "2002156_07.jpg", + "2002156_08.jpg", + "2002156_09.jpg", + "2002156_10.jpg", + "2002156_11.jpg", + "2002156_12.jpg", + "2002156_13.jpg", + "2002156_14.jpg", + "2002156_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A love scenes from two Russian. Between pink trees. Full screen. First pirson. In ten seconds. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001269", + "images": [ + "6001269_00.jpg", + "6001269_01.jpg", + "6001269_02.jpg", + "6001269_03.jpg", + "6001269_04.jpg", + "6001269_05.jpg", + "6001269_06.jpg", + "6001269_07.jpg", + "6001269_08.jpg", + "6001269_09.jpg", + "6001269_10.jpg", + "6001269_11.jpg", + "6001269_12.jpg", + "6001269_13.jpg", + "6001269_14.jpg", + "6001269_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a giant lake with flowers seen from above, aerial view, realistic photo. Zoom in. Message: RACHEL (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "4000929", + "images": [ + "4000929_00.jpg", + "4000929_01.jpg", + "4000929_02.jpg", + "4000929_03.jpg", + "4000929_04.jpg", + "4000929_05.jpg", + "4000929_06.jpg", + "4000929_07.jpg", + "4000929_08.jpg", + "4000929_09.jpg", + "4000929_10.jpg", + "4000929_11.jpg", + "4000929_12.jpg", + "4000929_13.jpg", + "4000929_14.jpg", + "4000929_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"concept art style of a colorful and vibrant forest \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "8000454", + "images": [ + "8000454_00.jpg", + "8000454_01.jpg", + "8000454_02.jpg", + "8000454_03.jpg", + "8000454_04.jpg", + "8000454_05.jpg", + "8000454_06.jpg", + "8000454_07.jpg", + "8000454_08.jpg", + "8000454_09.jpg", + "8000454_10.jpg", + "8000454_11.jpg", + "8000454_12.jpg", + "8000454_13.jpg", + "8000454_14.jpg", + "8000454_15.jpg", + "8000454_16.jpg", + "8000454_17.jpg", + "8000454_18.jpg", + "8000454_19.jpg", + "8000454_20.jpg", + "8000454_21.jpg", + "8000454_22.jpg", + "8000454_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Every day must be Sunday Amusement park inside the school \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001748", + "images": [ + "4001748_00.jpg", + "4001748_01.jpg", + "4001748_02.jpg", + "4001748_03.jpg", + "4001748_04.jpg", + "4001748_05.jpg", + "4001748_06.jpg", + "4001748_07.jpg", + "4001748_08.jpg", + "4001748_09.jpg", + "4001748_10.jpg", + "4001748_11.jpg", + "4001748_12.jpg", + "4001748_13.jpg", + "4001748_14.jpg", + "4001748_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The mother is seen sweeping the floors, dusting furniture, and arranging the house. Make the atmosphere calm but mysterious. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002979", + "images": [ + "3002979_00.jpg", + "3002979_01.jpg", + "3002979_02.jpg", + "3002979_03.jpg", + "3002979_04.jpg", + "3002979_05.jpg", + "3002979_06.jpg", + "3002979_07.jpg", + "3002979_08.jpg", + "3002979_09.jpg", + "3002979_10.jpg", + "3002979_11.jpg", + "3002979_12.jpg", + "3002979_13.jpg", + "3002979_14.jpg", + "3002979_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \": A quiet suburban neighborhood in Hounslow, London, showing a young Navinder Singh Sarao with his family \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "4000997", + "images": [ + "4000997_00.jpg", + "4000997_01.jpg", + "4000997_02.jpg", + "4000997_03.jpg", + "4000997_04.jpg", + "4000997_05.jpg", + "4000997_06.jpg", + "4000997_07.jpg", + "4000997_08.jpg", + "4000997_09.jpg", + "4000997_10.jpg", + "4000997_11.jpg", + "4000997_12.jpg", + "4000997_13.jpg", + "4000997_14.jpg", + "4000997_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"whale in space near an antimatter star, giant space squid emerges from star \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "1001035", + "images": [ + "1001035_00.jpg", + "1001035_01.jpg", + "1001035_02.jpg", + "1001035_03.jpg", + "1001035_04.jpg", + "1001035_05.jpg", + "1001035_06.jpg", + "1001035_07.jpg", + "1001035_08.jpg", + "1001035_09.jpg", + "1001035_10.jpg", + "1001035_11.jpg", + "1001035_12.jpg", + "1001035_13.jpg", + "1001035_14.jpg", + "1001035_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"show crypto market in tanking and add sound track. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000446", + "images": [ + "3000446_00.jpg", + "3000446_01.jpg", + "3000446_02.jpg", + "3000446_03.jpg", + "3000446_04.jpg", + "3000446_05.jpg", + "3000446_06.jpg", + "3000446_07.jpg", + "3000446_08.jpg", + "3000446_09.jpg", + "3000446_10.jpg", + "3000446_11.jpg", + "3000446_12.jpg", + "3000446_13.jpg", + "3000446_14.jpg", + "3000446_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"3 Beach Buggies driving at beach, blue sky, realistic, hd, cinematic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "8000267", + "images": [ + "8000267_00.jpg", + "8000267_01.jpg", + "8000267_02.jpg", + "8000267_03.jpg", + "8000267_04.jpg", + "8000267_05.jpg", + "8000267_06.jpg", + "8000267_07.jpg", + "8000267_08.jpg", + "8000267_09.jpg", + "8000267_10.jpg", + "8000267_11.jpg", + "8000267_12.jpg", + "8000267_13.jpg", + "8000267_14.jpg", + "8000267_15.jpg", + "8000267_16.jpg", + "8000267_17.jpg", + "8000267_18.jpg", + "8000267_19.jpg", + "8000267_20.jpg", + "8000267_21.jpg", + "8000267_22.jpg", + "8000267_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A hand holding a broken chain, liberation, cinematic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "5000995", + "images": [ + "5000995_00.jpg", + "5000995_01.jpg", + "5000995_02.jpg", + "5000995_03.jpg", + "5000995_04.jpg", + "5000995_05.jpg", + "5000995_06.jpg", + "5000995_07.jpg", + "5000995_08.jpg", + "5000995_09.jpg", + "5000995_10.jpg", + "5000995_11.jpg", + "5000995_12.jpg", + "5000995_13.jpg", + "5000995_14.jpg", + "5000995_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"One misty evening, Detective Evelyn Harper received an anonymous letter at the Ravenswood Police Department \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "7001407", + "images": [ + "7001407_00.jpg", + "7001407_01.jpg", + "7001407_02.jpg", + "7001407_03.jpg", + "7001407_04.jpg", + "7001407_05.jpg", + "7001407_06.jpg", + "7001407_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"jiafei sings and her hair is blown by wind background coconut beach blue lagoon and dragon \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002472", + "images": [ + "3002472_00.jpg", + "3002472_01.jpg", + "3002472_02.jpg", + "3002472_03.jpg", + "3002472_04.jpg", + "3002472_05.jpg", + "3002472_06.jpg", + "3002472_07.jpg", + "3002472_08.jpg", + "3002472_09.jpg", + "3002472_10.jpg", + "3002472_11.jpg", + "3002472_12.jpg", + "3002472_13.jpg", + "3002472_14.jpg", + "3002472_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a video of a racing car jumping over a building \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "2001266", + "images": [ + "2001266_00.jpg", + "2001266_01.jpg", + "2001266_02.jpg", + "2001266_03.jpg", + "2001266_04.jpg", + "2001266_05.jpg", + "2001266_06.jpg", + "2001266_07.jpg", + "2001266_08.jpg", + "2001266_09.jpg", + "2001266_10.jpg", + "2001266_11.jpg", + "2001266_12.jpg", + "2001266_13.jpg", + "2001266_14.jpg", + "2001266_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Hikers are drinking coffee at the top of the mountain \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000326", + "images": [ + "4000326_00.jpg", + "4000326_01.jpg", + "4000326_02.jpg", + "4000326_03.jpg", + "4000326_04.jpg", + "4000326_05.jpg", + "4000326_06.jpg", + "4000326_07.jpg", + "4000326_08.jpg", + "4000326_09.jpg", + "4000326_10.jpg", + "4000326_11.jpg", + "4000326_12.jpg", + "4000326_13.jpg", + "4000326_14.jpg", + "4000326_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"darth vader, disappointed watching the news, ultra 4k, hdr, apple video \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2000088", + "images": [ + "2000088_00.jpg", + "2000088_01.jpg", + "2000088_02.jpg", + "2000088_03.jpg", + "2000088_04.jpg", + "2000088_05.jpg", + "2000088_06.jpg", + "2000088_07.jpg", + "2000088_08.jpg", + "2000088_09.jpg", + "2000088_10.jpg", + "2000088_11.jpg", + "2000088_12.jpg", + "2000088_13.jpg", + "2000088_14.jpg", + "2000088_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"god of destruction walking ,anime style ,4k,HD quality \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "6000932", + "images": [ + "6000932_00.jpg", + "6000932_01.jpg", + "6000932_02.jpg", + "6000932_03.jpg", + "6000932_04.jpg", + "6000932_05.jpg", + "6000932_06.jpg", + "6000932_07.jpg", + "6000932_08.jpg", + "6000932_09.jpg", + "6000932_10.jpg", + "6000932_11.jpg", + "6000932_12.jpg", + "6000932_13.jpg", + "6000932_14.jpg", + "6000932_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"raw, 4k, realistic, girl, walking, rain, street, cars, \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000693", + "images": [ + "6000693_00.jpg", + "6000693_01.jpg", + "6000693_02.jpg", + "6000693_03.jpg", + "6000693_04.jpg", + "6000693_05.jpg", + "6000693_06.jpg", + "6000693_07.jpg", + "6000693_08.jpg", + "6000693_09.jpg", + "6000693_10.jpg", + "6000693_11.jpg", + "6000693_12.jpg", + "6000693_13.jpg", + "6000693_14.jpg", + "6000693_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Daniel in the den of angry lions but along with Daniel there is an angel of light \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001412", + "images": [ + "4001412_00.jpg", + "4001412_01.jpg", + "4001412_02.jpg", + "4001412_03.jpg", + "4001412_04.jpg", + "4001412_05.jpg", + "4001412_06.jpg", + "4001412_07.jpg", + "4001412_08.jpg", + "4001412_09.jpg", + "4001412_10.jpg", + "4001412_11.jpg", + "4001412_12.jpg", + "4001412_13.jpg", + "4001412_14.jpg", + "4001412_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"dream painting small boy look at the christmas food which on the table \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000343", + "images": [ + "6000343_00.jpg", + "6000343_01.jpg", + "6000343_02.jpg", + "6000343_03.jpg", + "6000343_04.jpg", + "6000343_05.jpg", + "6000343_06.jpg", + "6000343_07.jpg", + "6000343_08.jpg", + "6000343_09.jpg", + "6000343_10.jpg", + "6000343_11.jpg", + "6000343_12.jpg", + "6000343_13.jpg", + "6000343_14.jpg", + "6000343_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A Young Girl with black hair and her hair is flow \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001468", + "images": [ + "6001468_00.jpg", + "6001468_01.jpg", + "6001468_02.jpg", + "6001468_03.jpg", + "6001468_04.jpg", + "6001468_05.jpg", + "6001468_06.jpg", + "6001468_07.jpg", + "6001468_08.jpg", + "6001468_09.jpg", + "6001468_10.jpg", + "6001468_11.jpg", + "6001468_12.jpg", + "6001468_13.jpg", + "6001468_14.jpg", + "6001468_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"2d kawaii logo cute bright store picture 3d textures purple black pink blue \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000890", + "images": [ + "6000890_00.jpg", + "6000890_01.jpg", + "6000890_02.jpg", + "6000890_03.jpg", + "6000890_04.jpg", + "6000890_05.jpg", + "6000890_06.jpg", + "6000890_07.jpg", + "6000890_08.jpg", + "6000890_09.jpg", + "6000890_10.jpg", + "6000890_11.jpg", + "6000890_12.jpg", + "6000890_13.jpg", + "6000890_14.jpg", + "6000890_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"the best cherries of the season \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000532", + "images": [ + "3000532_00.jpg", + "3000532_01.jpg", + "3000532_02.jpg", + "3000532_03.jpg", + "3000532_04.jpg", + "3000532_05.jpg", + "3000532_06.jpg", + "3000532_07.jpg", + "3000532_08.jpg", + "3000532_09.jpg", + "3000532_10.jpg", + "3000532_11.jpg", + "3000532_12.jpg", + "3000532_13.jpg", + "3000532_14.jpg", + "3000532_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"take me to your leader, stop motion animation, claymation, 1950s animation \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "6001069", + "images": [ + "6001069_00.jpg", + "6001069_01.jpg", + "6001069_02.jpg", + "6001069_03.jpg", + "6001069_04.jpg", + "6001069_05.jpg", + "6001069_06.jpg", + "6001069_07.jpg", + "6001069_08.jpg", + "6001069_09.jpg", + "6001069_10.jpg", + "6001069_11.jpg", + "6001069_12.jpg", + "6001069_13.jpg", + "6001069_14.jpg", + "6001069_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"They developed new tools and techniques that enabled them to build complex structures, such as their famous ziggurats \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "5000865", + "images": [ + "5000865_00.jpg", + "5000865_01.jpg", + "5000865_02.jpg", + "5000865_03.jpg", + "5000865_04.jpg", + "5000865_05.jpg", + "5000865_06.jpg", + "5000865_07.jpg", + "5000865_08.jpg", + "5000865_09.jpg", + "5000865_10.jpg", + "5000865_11.jpg", + "5000865_12.jpg", + "5000865_13.jpg", + "5000865_14.jpg", + "5000865_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"baking together and making a mess with flour everywhere,pixar style,4k , Cartoon \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "0002196", + "images": [ + "0002196_00.jpg", + "0002196_01.jpg", + "0002196_02.jpg", + "0002196_03.jpg", + "0002196_04.jpg", + "0002196_05.jpg", + "0002196_06.jpg", + "0002196_07.jpg", + "0002196_08.jpg", + "0002196_09.jpg", + "0002196_10.jpg", + "0002196_11.jpg", + "0002196_12.jpg", + "0002196_13.jpg", + "0002196_14.jpg", + "0002196_15.jpg", + "0002196_16.jpg", + "0002196_17.jpg", + "0002196_18.jpg", + "0002196_19.jpg", + "0002196_20.jpg", + "0002196_21.jpg", + "0002196_22.jpg", + "0002196_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Add ambient sounds of ancient creatures and rustling leaves to set the eerie atmosphere \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001383", + "images": [ + "6001383_00.jpg", + "6001383_01.jpg", + "6001383_02.jpg", + "6001383_03.jpg", + "6001383_04.jpg", + "6001383_05.jpg", + "6001383_06.jpg", + "6001383_07.jpg", + "6001383_08.jpg", + "6001383_09.jpg", + "6001383_10.jpg", + "6001383_11.jpg", + "6001383_12.jpg", + "6001383_13.jpg", + "6001383_14.jpg", + "6001383_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Sasquatch is afraid of conspiracy theories \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001899", + "images": [ + "4001899_00.jpg", + "4001899_01.jpg", + "4001899_02.jpg", + "4001899_03.jpg", + "4001899_04.jpg", + "4001899_05.jpg", + "4001899_06.jpg", + "4001899_07.jpg", + "4001899_08.jpg", + "4001899_09.jpg", + "4001899_10.jpg", + "4001899_11.jpg", + "4001899_12.jpg", + "4001899_13.jpg", + "4001899_14.jpg", + "4001899_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"now transformed into a living testament of filial devotion, \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "3001514", + "images": [ + "3001514_00.jpg", + "3001514_01.jpg", + "3001514_02.jpg", + "3001514_03.jpg", + "3001514_04.jpg", + "3001514_05.jpg", + "3001514_06.jpg", + "3001514_07.jpg", + "3001514_08.jpg", + "3001514_09.jpg", + "3001514_10.jpg", + "3001514_11.jpg", + "3001514_12.jpg", + "3001514_13.jpg", + "3001514_14.jpg", + "3001514_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"surprised boys look at the ceiling \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "0000037", + "images": [ + "0000037_00.jpg", + "0000037_01.jpg", + "0000037_02.jpg", + "0000037_03.jpg", + "0000037_04.jpg", + "0000037_05.jpg", + "0000037_06.jpg", + "0000037_07.jpg", + "0000037_08.jpg", + "0000037_09.jpg", + "0000037_10.jpg", + "0000037_11.jpg", + "0000037_12.jpg", + "0000037_13.jpg", + "0000037_14.jpg", + "0000037_15.jpg", + "0000037_16.jpg", + "0000037_17.jpg", + "0000037_18.jpg", + "0000037_19.jpg", + "0000037_20.jpg", + "0000037_21.jpg", + "0000037_22.jpg", + "0000037_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"George W. Bush speaking to Walter White, Christmas, snowing, Breaking Bad, 2012 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002369", + "images": [ + "1002369_00.jpg", + "1002369_01.jpg", + "1002369_02.jpg", + "1002369_03.jpg", + "1002369_04.jpg", + "1002369_05.jpg", + "1002369_06.jpg", + "1002369_07.jpg", + "1002369_08.jpg", + "1002369_09.jpg", + "1002369_10.jpg", + "1002369_11.jpg", + "1002369_12.jpg", + "1002369_13.jpg", + "1002369_14.jpg", + "1002369_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A cute dog but getting startled by a loud noise and jumping into a bowl of popcorn, 8k, hyper detailed, highly detailed \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001162", + "images": [ + "6001162_00.jpg", + "6001162_01.jpg", + "6001162_02.jpg", + "6001162_03.jpg", + "6001162_04.jpg", + "6001162_05.jpg", + "6001162_06.jpg", + "6001162_07.jpg", + "6001162_08.jpg", + "6001162_09.jpg", + "6001162_10.jpg", + "6001162_11.jpg", + "6001162_12.jpg", + "6001162_13.jpg", + "6001162_14.jpg", + "6001162_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A primary school student with a worried expression \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "0002431", + "images": [ + "0002431_00.jpg", + "0002431_01.jpg", + "0002431_02.jpg", + "0002431_03.jpg", + "0002431_04.jpg", + "0002431_05.jpg", + "0002431_06.jpg", + "0002431_07.jpg", + "0002431_08.jpg", + "0002431_09.jpg", + "0002431_10.jpg", + "0002431_11.jpg", + "0002431_12.jpg", + "0002431_13.jpg", + "0002431_14.jpg", + "0002431_15.jpg", + "0002431_16.jpg", + "0002431_17.jpg", + "0002431_18.jpg", + "0002431_19.jpg", + "0002431_20.jpg", + "0002431_21.jpg", + "0002431_22.jpg", + "0002431_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a colorful image of 2 people walking on a beach watching high winds and high tides rolling in. Brilliant \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002436", + "images": [ + "1002436_00.jpg", + "1002436_01.jpg", + "1002436_02.jpg", + "1002436_03.jpg", + "1002436_04.jpg", + "1002436_05.jpg", + "1002436_06.jpg", + "1002436_07.jpg", + "1002436_08.jpg", + "1002436_09.jpg", + "1002436_10.jpg", + "1002436_11.jpg", + "1002436_12.jpg", + "1002436_13.jpg", + "1002436_14.jpg", + "1002436_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"live man floating and flying through the air \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002708", + "images": [ + "2002708_00.jpg", + "2002708_01.jpg", + "2002708_02.jpg", + "2002708_03.jpg", + "2002708_04.jpg", + "2002708_05.jpg", + "2002708_06.jpg", + "2002708_07.jpg", + "2002708_08.jpg", + "2002708_09.jpg", + "2002708_10.jpg", + "2002708_11.jpg", + "2002708_12.jpg", + "2002708_13.jpg", + "2002708_14.jpg", + "2002708_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Make images of Indian scout motorcycles, each one of a different color, the first will be green, the second will be white and the third will be red with a road background. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001056", + "images": [ + "7001056_00.jpg", + "7001056_01.jpg", + "7001056_02.jpg", + "7001056_03.jpg", + "7001056_04.jpg", + "7001056_05.jpg", + "7001056_06.jpg", + "7001056_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a blond hair and green eyes warrior lady on a walking dark horse \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2000791", + "images": [ + "2000791_00.jpg", + "2000791_01.jpg", + "2000791_02.jpg", + "2000791_03.jpg", + "2000791_04.jpg", + "2000791_05.jpg", + "2000791_06.jpg", + "2000791_07.jpg", + "2000791_08.jpg", + "2000791_09.jpg", + "2000791_10.jpg", + "2000791_11.jpg", + "2000791_12.jpg", + "2000791_13.jpg", + "2000791_14.jpg", + "2000791_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \": pine tree mountain top . Raining with thunderstorm . Panning camera \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000228", + "images": [ + "6000228_00.jpg", + "6000228_01.jpg", + "6000228_02.jpg", + "6000228_03.jpg", + "6000228_04.jpg", + "6000228_05.jpg", + "6000228_06.jpg", + "6000228_07.jpg", + "6000228_08.jpg", + "6000228_09.jpg", + "6000228_10.jpg", + "6000228_11.jpg", + "6000228_12.jpg", + "6000228_13.jpg", + "6000228_14.jpg", + "6000228_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"night sky with ufo invasion shining light on the city \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001880", + "images": [ + "6001880_00.jpg", + "6001880_01.jpg", + "6001880_02.jpg", + "6001880_03.jpg", + "6001880_04.jpg", + "6001880_05.jpg", + "6001880_06.jpg", + "6001880_07.jpg", + "6001880_08.jpg", + "6001880_09.jpg", + "6001880_10.jpg", + "6001880_11.jpg", + "6001880_12.jpg", + "6001880_13.jpg", + "6001880_14.jpg", + "6001880_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The two hugged each other, their clothes flying gently in the wind \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001169", + "images": [ + "4001169_00.jpg", + "4001169_01.jpg", + "4001169_02.jpg", + "4001169_03.jpg", + "4001169_04.jpg", + "4001169_05.jpg", + "4001169_06.jpg", + "4001169_07.jpg", + "4001169_08.jpg", + "4001169_09.jpg", + "4001169_10.jpg", + "4001169_11.jpg", + "4001169_12.jpg", + "4001169_13.jpg", + "4001169_14.jpg", + "4001169_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a mini studio for a cricket fan with team india collectives, podcast studio, indian team, cricket studio, commentary studio, realistic setting, \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2000620", + "images": [ + "2000620_00.jpg", + "2000620_01.jpg", + "2000620_02.jpg", + "2000620_03.jpg", + "2000620_04.jpg", + "2000620_05.jpg", + "2000620_06.jpg", + "2000620_07.jpg", + "2000620_08.jpg", + "2000620_09.jpg", + "2000620_10.jpg", + "2000620_11.jpg", + "2000620_12.jpg", + "2000620_13.jpg", + "2000620_14.jpg", + "2000620_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"titan musclare black robot in suit and purple metal lightning core, big old tv head and cannon blasters in arms staying on the city \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "0001900", + "images": [ + "0001900_00.jpg", + "0001900_01.jpg", + "0001900_02.jpg", + "0001900_03.jpg", + "0001900_04.jpg", + "0001900_05.jpg", + "0001900_06.jpg", + "0001900_07.jpg", + "0001900_08.jpg", + "0001900_09.jpg", + "0001900_10.jpg", + "0001900_11.jpg", + "0001900_12.jpg", + "0001900_13.jpg", + "0001900_14.jpg", + "0001900_15.jpg", + "0001900_16.jpg", + "0001900_17.jpg", + "0001900_18.jpg", + "0001900_19.jpg", + "0001900_20.jpg", + "0001900_21.jpg", + "0001900_22.jpg", + "0001900_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"There are conchs crawling under the sea. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001356", + "images": [ + "1001356_00.jpg", + "1001356_01.jpg", + "1001356_02.jpg", + "1001356_03.jpg", + "1001356_04.jpg", + "1001356_05.jpg", + "1001356_06.jpg", + "1001356_07.jpg", + "1001356_08.jpg", + "1001356_09.jpg", + "1001356_10.jpg", + "1001356_11.jpg", + "1001356_12.jpg", + "1001356_13.jpg", + "1001356_14.jpg", + "1001356_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a woman driving a car on a mountain \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3001117", + "images": [ + "3001117_00.jpg", + "3001117_01.jpg", + "3001117_02.jpg", + "3001117_03.jpg", + "3001117_04.jpg", + "3001117_05.jpg", + "3001117_06.jpg", + "3001117_07.jpg", + "3001117_08.jpg", + "3001117_09.jpg", + "3001117_10.jpg", + "3001117_11.jpg", + "3001117_12.jpg", + "3001117_13.jpg", + "3001117_14.jpg", + "3001117_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"different pros babysitter,handyman, tutor, photographer, personal trainer are shown on the map of a city \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "4001075", + "images": [ + "4001075_00.jpg", + "4001075_01.jpg", + "4001075_02.jpg", + "4001075_03.jpg", + "4001075_04.jpg", + "4001075_05.jpg", + "4001075_06.jpg", + "4001075_07.jpg", + "4001075_08.jpg", + "4001075_09.jpg", + "4001075_10.jpg", + "4001075_11.jpg", + "4001075_12.jpg", + "4001075_13.jpg", + "4001075_14.jpg", + "4001075_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"May the hearts of those who witness this vision be filled with reverence and gratitude, for they stand before GOD himself, the architect of the universe, the conductor of the cosmic symphony, the light that illuminates every corner of existence, cinematic style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002855", + "images": [ + "3002855_00.jpg", + "3002855_01.jpg", + "3002855_02.jpg", + "3002855_03.jpg", + "3002855_04.jpg", + "3002855_05.jpg", + "3002855_06.jpg", + "3002855_07.jpg", + "3002855_08.jpg", + "3002855_09.jpg", + "3002855_10.jpg", + "3002855_11.jpg", + "3002855_12.jpg", + "3002855_13.jpg", + "3002855_14.jpg", + "3002855_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a 50 years old man wearing faded blue long sleeves shirt and long sleeves faded jeans walking in the jungle in the early morning, dim sun light, photorealistic images, \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000237", + "images": [ + "7000237_00.jpg", + "7000237_01.jpg", + "7000237_02.jpg", + "7000237_03.jpg", + "7000237_04.jpg", + "7000237_05.jpg", + "7000237_06.jpg", + "7000237_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a close up of a boy with purple eyes opening his eyes, anime style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002107", + "images": [ + "2002107_00.jpg", + "2002107_01.jpg", + "2002107_02.jpg", + "2002107_03.jpg", + "2002107_04.jpg", + "2002107_05.jpg", + "2002107_06.jpg", + "2002107_07.jpg", + "2002107_08.jpg", + "2002107_09.jpg", + "2002107_10.jpg", + "2002107_11.jpg", + "2002107_12.jpg", + "2002107_13.jpg", + "2002107_14.jpg", + "2002107_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"realistic photo of terminator robots in an apocalypsis world, fire, smoke, detailed, 4k Message: genesis (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000455", + "images": [ + "4000455_00.jpg", + "4000455_01.jpg", + "4000455_02.jpg", + "4000455_03.jpg", + "4000455_04.jpg", + "4000455_05.jpg", + "4000455_06.jpg", + "4000455_07.jpg", + "4000455_08.jpg", + "4000455_09.jpg", + "4000455_10.jpg", + "4000455_11.jpg", + "4000455_12.jpg", + "4000455_13.jpg", + "4000455_14.jpg", + "4000455_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"fantastic lights from the vulcano, uhd, ultradetailed \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3001999", + "images": [ + "3001999_00.jpg", + "3001999_01.jpg", + "3001999_02.jpg", + "3001999_03.jpg", + "3001999_04.jpg", + "3001999_05.jpg", + "3001999_06.jpg", + "3001999_07.jpg", + "3001999_08.jpg", + "3001999_09.jpg", + "3001999_10.jpg", + "3001999_11.jpg", + "3001999_12.jpg", + "3001999_13.jpg", + "3001999_14.jpg", + "3001999_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Generate 10 second cartoon of a character expressing intense emotions: joy, anger, fear, surprise. Bright, exaggerated style. Character gestures wildly. Simple background. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "6000051", + "images": [ + "6000051_00.jpg", + "6000051_01.jpg", + "6000051_02.jpg", + "6000051_03.jpg", + "6000051_04.jpg", + "6000051_05.jpg", + "6000051_06.jpg", + "6000051_07.jpg", + "6000051_08.jpg", + "6000051_09.jpg", + "6000051_10.jpg", + "6000051_11.jpg", + "6000051_12.jpg", + "6000051_13.jpg", + "6000051_14.jpg", + "6000051_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Anime girl light blue hair pink clothes \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "1002083", + "images": [ + "1002083_00.jpg", + "1002083_01.jpg", + "1002083_02.jpg", + "1002083_03.jpg", + "1002083_04.jpg", + "1002083_05.jpg", + "1002083_06.jpg", + "1002083_07.jpg", + "1002083_08.jpg", + "1002083_09.jpg", + "1002083_10.jpg", + "1002083_11.jpg", + "1002083_12.jpg", + "1002083_13.jpg", + "1002083_14.jpg", + "1002083_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"short film, realistic, human, 16:9, dark, art house, forest, sacrificing, burning, forest background, real, children, worshipping, flames, story telling, multiple scenes, grainy, raining, horror, multiple images of sadness, crying, devil worship, film, movie \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1002899", + "images": [ + "1002899_00.jpg", + "1002899_01.jpg", + "1002899_02.jpg", + "1002899_03.jpg", + "1002899_04.jpg", + "1002899_05.jpg", + "1002899_06.jpg", + "1002899_07.jpg", + "1002899_08.jpg", + "1002899_09.jpg", + "1002899_10.jpg", + "1002899_11.jpg", + "1002899_12.jpg", + "1002899_13.jpg", + "1002899_14.jpg", + "1002899_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"image for concentration when studying,futuristic image \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001518", + "images": [ + "4001518_00.jpg", + "4001518_01.jpg", + "4001518_02.jpg", + "4001518_03.jpg", + "4001518_04.jpg", + "4001518_05.jpg", + "4001518_06.jpg", + "4001518_07.jpg", + "4001518_08.jpg", + "4001518_09.jpg", + "4001518_10.jpg", + "4001518_11.jpg", + "4001518_12.jpg", + "4001518_13.jpg", + "4001518_14.jpg", + "4001518_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The electrician is burning struck by lightning and moaning in unbearable pain. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000768", + "images": [ + "6000768_00.jpg", + "6000768_01.jpg", + "6000768_02.jpg", + "6000768_03.jpg", + "6000768_04.jpg", + "6000768_05.jpg", + "6000768_06.jpg", + "6000768_07.jpg", + "6000768_08.jpg", + "6000768_09.jpg", + "6000768_10.jpg", + "6000768_11.jpg", + "6000768_12.jpg", + "6000768_13.jpg", + "6000768_14.jpg", + "6000768_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a village boy running in gree effect \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3001363", + "images": [ + "3001363_00.jpg", + "3001363_01.jpg", + "3001363_02.jpg", + "3001363_03.jpg", + "3001363_04.jpg", + "3001363_05.jpg", + "3001363_06.jpg", + "3001363_07.jpg", + "3001363_08.jpg", + "3001363_09.jpg", + "3001363_10.jpg", + "3001363_11.jpg", + "3001363_12.jpg", + "3001363_13.jpg", + "3001363_14.jpg", + "3001363_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Show the child cautiously exploring the dimly lit interior of the house. Use shadows to hint at the presence of supernatural entities or eerie happenings, such as a ghostly figure or a swinging old swing that moves on its own. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "0001342", + "images": [ + "0001342_00.jpg", + "0001342_01.jpg", + "0001342_02.jpg", + "0001342_03.jpg", + "0001342_04.jpg", + "0001342_05.jpg", + "0001342_06.jpg", + "0001342_07.jpg", + "0001342_08.jpg", + "0001342_09.jpg", + "0001342_10.jpg", + "0001342_11.jpg", + "0001342_12.jpg", + "0001342_13.jpg", + "0001342_14.jpg", + "0001342_15.jpg", + "0001342_16.jpg", + "0001342_17.jpg", + "0001342_18.jpg", + "0001342_19.jpg", + "0001342_20.jpg", + "0001342_21.jpg", + "0001342_22.jpg", + "0001342_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"the first Sparkling Stone revealed itself, emitting a soft glow. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "8000783", + "images": [ + "8000783_00.jpg", + "8000783_01.jpg", + "8000783_02.jpg", + "8000783_03.jpg", + "8000783_04.jpg", + "8000783_05.jpg", + "8000783_06.jpg", + "8000783_07.jpg", + "8000783_08.jpg", + "8000783_09.jpg", + "8000783_10.jpg", + "8000783_11.jpg", + "8000783_12.jpg", + "8000783_13.jpg", + "8000783_14.jpg", + "8000783_15.jpg", + "8000783_16.jpg", + "8000783_17.jpg", + "8000783_18.jpg", + "8000783_19.jpg", + "8000783_20.jpg", + "8000783_21.jpg", + "8000783_22.jpg", + "8000783_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The magical meadow not only transforms the land but also the young man, who learns that true magic lies in the power of friendship, love and courage. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1002669", + "images": [ + "1002669_00.jpg", + "1002669_01.jpg", + "1002669_02.jpg", + "1002669_03.jpg", + "1002669_04.jpg", + "1002669_05.jpg", + "1002669_06.jpg", + "1002669_07.jpg", + "1002669_08.jpg", + "1002669_09.jpg", + "1002669_10.jpg", + "1002669_11.jpg", + "1002669_12.jpg", + "1002669_13.jpg", + "1002669_14.jpg", + "1002669_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"make a presentation of a hinged rocking chair in a natural style with green plants and exotic birds in make a presentation of a hinged rocking chair in a natural style with green plants and exotic birds inmake a presentation of a hinged rocking chair in a natural style with green plants and exotic birds \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000707", + "images": [ + "6000707_00.jpg", + "6000707_01.jpg", + "6000707_02.jpg", + "6000707_03.jpg", + "6000707_04.jpg", + "6000707_05.jpg", + "6000707_06.jpg", + "6000707_07.jpg", + "6000707_08.jpg", + "6000707_09.jpg", + "6000707_10.jpg", + "6000707_11.jpg", + "6000707_12.jpg", + "6000707_13.jpg", + "6000707_14.jpg", + "6000707_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"rays of light leaving the sun and making their way towards earth \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "8000515", + "images": [ + "8000515_00.jpg", + "8000515_01.jpg", + "8000515_02.jpg", + "8000515_03.jpg", + "8000515_04.jpg", + "8000515_05.jpg", + "8000515_06.jpg", + "8000515_07.jpg", + "8000515_08.jpg", + "8000515_09.jpg", + "8000515_10.jpg", + "8000515_11.jpg", + "8000515_12.jpg", + "8000515_13.jpg", + "8000515_14.jpg", + "8000515_15.jpg", + "8000515_16.jpg", + "8000515_17.jpg", + "8000515_18.jpg", + "8000515_19.jpg", + "8000515_20.jpg", + "8000515_21.jpg", + "8000515_22.jpg", + "8000515_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Jawan movie, shah rukh khan, stylish, cinematic footage, high quality \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "5000586", + "images": [ + "5000586_00.jpg", + "5000586_01.jpg", + "5000586_02.jpg", + "5000586_03.jpg", + "5000586_04.jpg", + "5000586_05.jpg", + "5000586_06.jpg", + "5000586_07.jpg", + "5000586_08.jpg", + "5000586_09.jpg", + "5000586_10.jpg", + "5000586_11.jpg", + "5000586_12.jpg", + "5000586_13.jpg", + "5000586_14.jpg", + "5000586_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"It is capable of speaking, writing and translating like humans. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "1000858", + "images": [ + "1000858_00.jpg", + "1000858_01.jpg", + "1000858_02.jpg", + "1000858_03.jpg", + "1000858_04.jpg", + "1000858_05.jpg", + "1000858_06.jpg", + "1000858_07.jpg", + "1000858_08.jpg", + "1000858_09.jpg", + "1000858_10.jpg", + "1000858_11.jpg", + "1000858_12.jpg", + "1000858_13.jpg", + "1000858_14.jpg", + "1000858_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"beyond time and space there is a construct that rotates \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0002110", + "images": [ + "0002110_00.jpg", + "0002110_01.jpg", + "0002110_02.jpg", + "0002110_03.jpg", + "0002110_04.jpg", + "0002110_05.jpg", + "0002110_06.jpg", + "0002110_07.jpg", + "0002110_08.jpg", + "0002110_09.jpg", + "0002110_10.jpg", + "0002110_11.jpg", + "0002110_12.jpg", + "0002110_13.jpg", + "0002110_14.jpg", + "0002110_15.jpg", + "0002110_16.jpg", + "0002110_17.jpg", + "0002110_18.jpg", + "0002110_19.jpg", + "0002110_20.jpg", + "0002110_21.jpg", + "0002110_22.jpg", + "0002110_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A small construction site with Eddie the Excavator in action, showcasing his powerful digging arm in motion, scooping up dirt. The background should show a partially dug earth area. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3001810", + "images": [ + "3001810_00.jpg", + "3001810_01.jpg", + "3001810_02.jpg", + "3001810_03.jpg", + "3001810_04.jpg", + "3001810_05.jpg", + "3001810_06.jpg", + "3001810_07.jpg", + "3001810_08.jpg", + "3001810_09.jpg", + "3001810_10.jpg", + "3001810_11.jpg", + "3001810_12.jpg", + "3001810_13.jpg", + "3001810_14.jpg", + "3001810_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"After this, the sounds of the palace and the forest also became quiet and people now find peace there. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "8000876", + "images": [ + "8000876_00.jpg", + "8000876_01.jpg", + "8000876_02.jpg", + "8000876_03.jpg", + "8000876_04.jpg", + "8000876_05.jpg", + "8000876_06.jpg", + "8000876_07.jpg", + "8000876_08.jpg", + "8000876_09.jpg", + "8000876_10.jpg", + "8000876_11.jpg", + "8000876_12.jpg", + "8000876_13.jpg", + "8000876_14.jpg", + "8000876_15.jpg", + "8000876_16.jpg", + "8000876_17.jpg", + "8000876_18.jpg", + "8000876_19.jpg", + "8000876_20.jpg", + "8000876_21.jpg", + "8000876_22.jpg", + "8000876_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a boy from a poor family and become a rich man \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000830", + "images": [ + "7000830_00.jpg", + "7000830_01.jpg", + "7000830_02.jpg", + "7000830_03.jpg", + "7000830_04.jpg", + "7000830_05.jpg", + "7000830_06.jpg", + "7000830_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a realistic picture of a drawn table, full of tools and colours Message: Pencil (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000562", + "images": [ + "4000562_00.jpg", + "4000562_01.jpg", + "4000562_02.jpg", + "4000562_03.jpg", + "4000562_04.jpg", + "4000562_05.jpg", + "4000562_06.jpg", + "4000562_07.jpg", + "4000562_08.jpg", + "4000562_09.jpg", + "4000562_10.jpg", + "4000562_11.jpg", + "4000562_12.jpg", + "4000562_13.jpg", + "4000562_14.jpg", + "4000562_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a soccer ball that robots can play with \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000317", + "images": [ + "3000317_00.jpg", + "3000317_01.jpg", + "3000317_02.jpg", + "3000317_03.jpg", + "3000317_04.jpg", + "3000317_05.jpg", + "3000317_06.jpg", + "3000317_07.jpg", + "3000317_08.jpg", + "3000317_09.jpg", + "3000317_10.jpg", + "3000317_11.jpg", + "3000317_12.jpg", + "3000317_13.jpg", + "3000317_14.jpg", + "3000317_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Craft a compelling prompt that captures the diverse colors, popular places, and social life of your village. Encourage the creation of a vivid and optimal depiction that showcases the prosperity and unity of your village \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3001521", + "images": [ + "3001521_00.jpg", + "3001521_01.jpg", + "3001521_02.jpg", + "3001521_03.jpg", + "3001521_04.jpg", + "3001521_05.jpg", + "3001521_06.jpg", + "3001521_07.jpg", + "3001521_08.jpg", + "3001521_09.jpg", + "3001521_10.jpg", + "3001521_11.jpg", + "3001521_12.jpg", + "3001521_13.jpg", + "3001521_14.jpg", + "3001521_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"guys with swords on horse in desert \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "4001136", + "images": [ + "4001136_00.jpg", + "4001136_01.jpg", + "4001136_02.jpg", + "4001136_03.jpg", + "4001136_04.jpg", + "4001136_05.jpg", + "4001136_06.jpg", + "4001136_07.jpg", + "4001136_08.jpg", + "4001136_09.jpg", + "4001136_10.jpg", + "4001136_11.jpg", + "4001136_12.jpg", + "4001136_13.jpg", + "4001136_14.jpg", + "4001136_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a royal palaces, Great Zimbabwe during the late Iron Age residences of the ruling elite, motion 4, camera panaramic right \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1000744", + "images": [ + "1000744_00.jpg", + "1000744_01.jpg", + "1000744_02.jpg", + "1000744_03.jpg", + "1000744_04.jpg", + "1000744_05.jpg", + "1000744_06.jpg", + "1000744_07.jpg", + "1000744_08.jpg", + "1000744_09.jpg", + "1000744_10.jpg", + "1000744_11.jpg", + "1000744_12.jpg", + "1000744_13.jpg", + "1000744_14.jpg", + "1000744_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a fish flying in the sky and wearing a tie \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1000471", + "images": [ + "1000471_00.jpg", + "1000471_01.jpg", + "1000471_02.jpg", + "1000471_03.jpg", + "1000471_04.jpg", + "1000471_05.jpg", + "1000471_06.jpg", + "1000471_07.jpg", + "1000471_08.jpg", + "1000471_09.jpg", + "1000471_10.jpg", + "1000471_11.jpg", + "1000471_12.jpg", + "1000471_13.jpg", + "1000471_14.jpg", + "1000471_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"polaroid of a person driving a car wearing a sun helmet \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1001015", + "images": [ + "1001015_00.jpg", + "1001015_01.jpg", + "1001015_02.jpg", + "1001015_03.jpg", + "1001015_04.jpg", + "1001015_05.jpg", + "1001015_06.jpg", + "1001015_07.jpg", + "1001015_08.jpg", + "1001015_09.jpg", + "1001015_10.jpg", + "1001015_11.jpg", + "1001015_12.jpg", + "1001015_13.jpg", + "1001015_14.jpg", + "1001015_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"An engineer is disassembling the car seats inside the vehicle, with wires and cables exposed and scattered around the scene. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0000255", + "images": [ + "0000255_00.jpg", + "0000255_01.jpg", + "0000255_02.jpg", + "0000255_03.jpg", + "0000255_04.jpg", + "0000255_05.jpg", + "0000255_06.jpg", + "0000255_07.jpg", + "0000255_08.jpg", + "0000255_09.jpg", + "0000255_10.jpg", + "0000255_11.jpg", + "0000255_12.jpg", + "0000255_13.jpg", + "0000255_14.jpg", + "0000255_15.jpg", + "0000255_16.jpg", + "0000255_17.jpg", + "0000255_18.jpg", + "0000255_19.jpg", + "0000255_20.jpg", + "0000255_21.jpg", + "0000255_22.jpg", + "0000255_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"During the confrontation, John takes aim and fires, initiating a fierce gun battle. The shot showcases the intense action, with bullets flying and sparks from ricochets, heightening the chaos \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0001078", + "images": [ + "0001078_00.jpg", + "0001078_01.jpg", + "0001078_02.jpg", + "0001078_03.jpg", + "0001078_04.jpg", + "0001078_05.jpg", + "0001078_06.jpg", + "0001078_07.jpg", + "0001078_08.jpg", + "0001078_09.jpg", + "0001078_10.jpg", + "0001078_11.jpg", + "0001078_12.jpg", + "0001078_13.jpg", + "0001078_14.jpg", + "0001078_15.jpg", + "0001078_16.jpg", + "0001078_17.jpg", + "0001078_18.jpg", + "0001078_19.jpg", + "0001078_20.jpg", + "0001078_21.jpg", + "0001078_22.jpg", + "0001078_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"create a digital child with dancing and enjoying snowfall \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4001757", + "images": [ + "4001757_00.jpg", + "4001757_01.jpg", + "4001757_02.jpg", + "4001757_03.jpg", + "4001757_04.jpg", + "4001757_05.jpg", + "4001757_06.jpg", + "4001757_07.jpg", + "4001757_08.jpg", + "4001757_09.jpg", + "4001757_10.jpg", + "4001757_11.jpg", + "4001757_12.jpg", + "4001757_13.jpg", + "4001757_14.jpg", + "4001757_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Alien standing in the jungle of an alien planet \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0000718", + "images": [ + "0000718_00.jpg", + "0000718_01.jpg", + "0000718_02.jpg", + "0000718_03.jpg", + "0000718_04.jpg", + "0000718_05.jpg", + "0000718_06.jpg", + "0000718_07.jpg", + "0000718_08.jpg", + "0000718_09.jpg", + "0000718_10.jpg", + "0000718_11.jpg", + "0000718_12.jpg", + "0000718_13.jpg", + "0000718_14.jpg", + "0000718_15.jpg", + "0000718_16.jpg", + "0000718_17.jpg", + "0000718_18.jpg", + "0000718_19.jpg", + "0000718_20.jpg", + "0000718_21.jpg", + "0000718_22.jpg", + "0000718_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"in an egyptian settings, create a man gazes out the window at people suffering from the mysterious ailment, their symptoms evident. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002229", + "images": [ + "2002229_00.jpg", + "2002229_01.jpg", + "2002229_02.jpg", + "2002229_03.jpg", + "2002229_04.jpg", + "2002229_05.jpg", + "2002229_06.jpg", + "2002229_07.jpg", + "2002229_08.jpg", + "2002229_09.jpg", + "2002229_10.jpg", + "2002229_11.jpg", + "2002229_12.jpg", + "2002229_13.jpg", + "2002229_14.jpg", + "2002229_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"growing a chair made of mushroom and bubble \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "4001321", + "images": [ + "4001321_00.jpg", + "4001321_01.jpg", + "4001321_02.jpg", + "4001321_03.jpg", + "4001321_04.jpg", + "4001321_05.jpg", + "4001321_06.jpg", + "4001321_07.jpg", + "4001321_08.jpg", + "4001321_09.jpg", + "4001321_10.jpg", + "4001321_11.jpg", + "4001321_12.jpg", + "4001321_13.jpg", + "4001321_14.jpg", + "4001321_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"In a dense forest, in a valley illuminated only by the pale light of stars, Karak, a hunter from the tribe of ancient people, prepares for a ritual. The red paint on his face sparkles against the backdrop of tongues of flame, creating a mysterious figure. Around him, in anticipation, stand ancient dinosaurs, their massive shadows dancing on the walls of the surrounding cliffs \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000289", + "images": [ + "5000289_00.jpg", + "5000289_01.jpg", + "5000289_02.jpg", + "5000289_03.jpg", + "5000289_04.jpg", + "5000289_05.jpg", + "5000289_06.jpg", + "5000289_07.jpg", + "5000289_08.jpg", + "5000289_09.jpg", + "5000289_10.jpg", + "5000289_11.jpg", + "5000289_12.jpg", + "5000289_13.jpg", + "5000289_14.jpg", + "5000289_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"nuclear explosion, masterpiece, movie, dynamic scene, action, hdr, 4k \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000475", + "images": [ + "3000475_00.jpg", + "3000475_01.jpg", + "3000475_02.jpg", + "3000475_03.jpg", + "3000475_04.jpg", + "3000475_05.jpg", + "3000475_06.jpg", + "3000475_07.jpg", + "3000475_08.jpg", + "3000475_09.jpg", + "3000475_10.jpg", + "3000475_11.jpg", + "3000475_12.jpg", + "3000475_13.jpg", + "3000475_14.jpg", + "3000475_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A beautiful woman happily walking through a city decorated with Christmas illuminations in the evening, anime style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "7000259", + "images": [ + "7000259_00.jpg", + "7000259_01.jpg", + "7000259_02.jpg", + "7000259_03.jpg", + "7000259_04.jpg", + "7000259_05.jpg", + "7000259_06.jpg", + "7000259_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A pair of people are passing through a transparent glass tunnel, with an overhead drone view of the mysterious, eerie atmosphere of the image \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000605", + "images": [ + "4000605_00.jpg", + "4000605_01.jpg", + "4000605_02.jpg", + "4000605_03.jpg", + "4000605_04.jpg", + "4000605_05.jpg", + "4000605_06.jpg", + "4000605_07.jpg", + "4000605_08.jpg", + "4000605_09.jpg", + "4000605_10.jpg", + "4000605_11.jpg", + "4000605_12.jpg", + "4000605_13.jpg", + "4000605_14.jpg", + "4000605_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a athletic man trains his Trizeps up and down, with a barbell in a fitness center for ten seconds \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "2001594", + "images": [ + "2001594_00.jpg", + "2001594_01.jpg", + "2001594_02.jpg", + "2001594_03.jpg", + "2001594_04.jpg", + "2001594_05.jpg", + "2001594_06.jpg", + "2001594_07.jpg", + "2001594_08.jpg", + "2001594_09.jpg", + "2001594_10.jpg", + "2001594_11.jpg", + "2001594_12.jpg", + "2001594_13.jpg", + "2001594_14.jpg", + "2001594_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"animated video of harry Potter doing magics \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "2002690", + "images": [ + "2002690_00.jpg", + "2002690_01.jpg", + "2002690_02.jpg", + "2002690_03.jpg", + "2002690_04.jpg", + "2002690_05.jpg", + "2002690_06.jpg", + "2002690_07.jpg", + "2002690_08.jpg", + "2002690_09.jpg", + "2002690_10.jpg", + "2002690_11.jpg", + "2002690_12.jpg", + "2002690_13.jpg", + "2002690_14.jpg", + "2002690_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"men worried about her aging face \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "0000755", + "images": [ + "0000755_00.jpg", + "0000755_01.jpg", + "0000755_02.jpg", + "0000755_03.jpg", + "0000755_04.jpg", + "0000755_05.jpg", + "0000755_06.jpg", + "0000755_07.jpg", + "0000755_08.jpg", + "0000755_09.jpg", + "0000755_10.jpg", + "0000755_11.jpg", + "0000755_12.jpg", + "0000755_13.jpg", + "0000755_14.jpg", + "0000755_15.jpg", + "0000755_16.jpg", + "0000755_17.jpg", + "0000755_18.jpg", + "0000755_19.jpg", + "0000755_20.jpg", + "0000755_21.jpg", + "0000755_22.jpg", + "0000755_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a raven flying over the battlefield \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000498", + "images": [ + "4000498_00.jpg", + "4000498_01.jpg", + "4000498_02.jpg", + "4000498_03.jpg", + "4000498_04.jpg", + "4000498_05.jpg", + "4000498_06.jpg", + "4000498_07.jpg", + "4000498_08.jpg", + "4000498_09.jpg", + "4000498_10.jpg", + "4000498_11.jpg", + "4000498_12.jpg", + "4000498_13.jpg", + "4000498_14.jpg", + "4000498_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The cucumber flag waving in the wind. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002447", + "images": [ + "1002447_00.jpg", + "1002447_01.jpg", + "1002447_02.jpg", + "1002447_03.jpg", + "1002447_04.jpg", + "1002447_05.jpg", + "1002447_06.jpg", + "1002447_07.jpg", + "1002447_08.jpg", + "1002447_09.jpg", + "1002447_10.jpg", + "1002447_11.jpg", + "1002447_12.jpg", + "1002447_13.jpg", + "1002447_14.jpg", + "1002447_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The cartoon opens with a picturesque view of a lively downtown, bustling with colorful characters and charming architecture \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0000339", + "images": [ + "0000339_00.jpg", + "0000339_01.jpg", + "0000339_02.jpg", + "0000339_03.jpg", + "0000339_04.jpg", + "0000339_05.jpg", + "0000339_06.jpg", + "0000339_07.jpg", + "0000339_08.jpg", + "0000339_09.jpg", + "0000339_10.jpg", + "0000339_11.jpg", + "0000339_12.jpg", + "0000339_13.jpg", + "0000339_14.jpg", + "0000339_15.jpg", + "0000339_16.jpg", + "0000339_17.jpg", + "0000339_18.jpg", + "0000339_19.jpg", + "0000339_20.jpg", + "0000339_21.jpg", + "0000339_22.jpg", + "0000339_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"super 8 footage of blimp flying over warzone \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 4\n temporal consistency: 4\n dynamic degree: 4\n text-to-video alignment: 4\n factual consistency: 4\n" + } + ] + }, + { + "id": "3000295", + "images": [ + "3000295_00.jpg", + "3000295_01.jpg", + "3000295_02.jpg", + "3000295_03.jpg", + "3000295_04.jpg", + "3000295_05.jpg", + "3000295_06.jpg", + "3000295_07.jpg", + "3000295_08.jpg", + "3000295_09.jpg", + "3000295_10.jpg", + "3000295_11.jpg", + "3000295_12.jpg", + "3000295_13.jpg", + "3000295_14.jpg", + "3000295_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"About the rain in urban Indonesia, and there was a child wearing a white shirt, black hair facing backwards in the urban area at night without wearing sandals and wearing shoes, he was standing there. Rain drops that make his body wet, 4k quality images and also HD ANIMATION ANIME STUDIOS \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "1002621", + "images": [ + "1002621_00.jpg", + "1002621_01.jpg", + "1002621_02.jpg", + "1002621_03.jpg", + "1002621_04.jpg", + "1002621_05.jpg", + "1002621_06.jpg", + "1002621_07.jpg", + "1002621_08.jpg", + "1002621_09.jpg", + "1002621_10.jpg", + "1002621_11.jpg", + "1002621_12.jpg", + "1002621_13.jpg", + "1002621_14.jpg", + "1002621_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a black range rover sport car driving on the road\nwith camera focusing on its tyer and rolling wheel \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "5000685", + "images": [ + "5000685_00.jpg", + "5000685_01.jpg", + "5000685_02.jpg", + "5000685_03.jpg", + "5000685_04.jpg", + "5000685_05.jpg", + "5000685_06.jpg", + "5000685_07.jpg", + "5000685_08.jpg", + "5000685_09.jpg", + "5000685_10.jpg", + "5000685_11.jpg", + "5000685_12.jpg", + "5000685_13.jpg", + "5000685_14.jpg", + "5000685_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a giant translucent angel covered in ancient gems flying over a burning city surrounded by beautiful glowing lights \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 2\n" + } + ] + }, + { + "id": "2000380", + "images": [ + "2000380_00.jpg", + "2000380_01.jpg", + "2000380_02.jpg", + "2000380_03.jpg", + "2000380_04.jpg", + "2000380_05.jpg", + "2000380_06.jpg", + "2000380_07.jpg", + "2000380_08.jpg", + "2000380_09.jpg", + "2000380_10.jpg", + "2000380_11.jpg", + "2000380_12.jpg", + "2000380_13.jpg", + "2000380_14.jpg", + "2000380_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A memorial statue of Abraham Lincoln sitting on a chair, with the words \u201cIn this temple, as in the hearts of the people for whom he saved the Union, the memory of Abraham Lincoln is enshrined forever\u201d \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000393", + "images": [ + "4000393_00.jpg", + "4000393_01.jpg", + "4000393_02.jpg", + "4000393_03.jpg", + "4000393_04.jpg", + "4000393_05.jpg", + "4000393_06.jpg", + "4000393_07.jpg", + "4000393_08.jpg", + "4000393_09.jpg", + "4000393_10.jpg", + "4000393_11.jpg", + "4000393_12.jpg", + "4000393_13.jpg", + "4000393_14.jpg", + "4000393_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"fabric in pink, fashion, india, realistic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002277", + "images": [ + "2002277_00.jpg", + "2002277_01.jpg", + "2002277_02.jpg", + "2002277_03.jpg", + "2002277_04.jpg", + "2002277_05.jpg", + "2002277_06.jpg", + "2002277_07.jpg", + "2002277_08.jpg", + "2002277_09.jpg", + "2002277_10.jpg", + "2002277_11.jpg", + "2002277_12.jpg", + "2002277_13.jpg", + "2002277_14.jpg", + "2002277_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A shot of the humble shoemaker, Sam, in his small workshop. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002894", + "images": [ + "3002894_00.jpg", + "3002894_01.jpg", + "3002894_02.jpg", + "3002894_03.jpg", + "3002894_04.jpg", + "3002894_05.jpg", + "3002894_06.jpg", + "3002894_07.jpg", + "3002894_08.jpg", + "3002894_09.jpg", + "3002894_10.jpg", + "3002894_11.jpg", + "3002894_12.jpg", + "3002894_13.jpg", + "3002894_14.jpg", + "3002894_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"mesmerizing depiction of a Pink Sand Beach akin to Plage aux Bahamas with moving waves \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "3002391", + "images": [ + "3002391_00.jpg", + "3002391_01.jpg", + "3002391_02.jpg", + "3002391_03.jpg", + "3002391_04.jpg", + "3002391_05.jpg", + "3002391_06.jpg", + "3002391_07.jpg", + "3002391_08.jpg", + "3002391_09.jpg", + "3002391_10.jpg", + "3002391_11.jpg", + "3002391_12.jpg", + "3002391_13.jpg", + "3002391_14.jpg", + "3002391_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"He devised an ingenious experiment.visibal footage, 4k, 8k. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002713", + "images": [ + "3002713_00.jpg", + "3002713_01.jpg", + "3002713_02.jpg", + "3002713_03.jpg", + "3002713_04.jpg", + "3002713_05.jpg", + "3002713_06.jpg", + "3002713_07.jpg", + "3002713_08.jpg", + "3002713_09.jpg", + "3002713_10.jpg", + "3002713_11.jpg", + "3002713_12.jpg", + "3002713_13.jpg", + "3002713_14.jpg", + "3002713_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"man be motivated and try change his life \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "2001184", + "images": [ + "2001184_00.jpg", + "2001184_01.jpg", + "2001184_02.jpg", + "2001184_03.jpg", + "2001184_04.jpg", + "2001184_05.jpg", + "2001184_06.jpg", + "2001184_07.jpg", + "2001184_08.jpg", + "2001184_09.jpg", + "2001184_10.jpg", + "2001184_11.jpg", + "2001184_12.jpg", + "2001184_13.jpg", + "2001184_14.jpg", + "2001184_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"samurai walking through puddles of ketchup in a cave where monsters in the elder scrolls style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "4001241", + "images": [ + "4001241_00.jpg", + "4001241_01.jpg", + "4001241_02.jpg", + "4001241_03.jpg", + "4001241_04.jpg", + "4001241_05.jpg", + "4001241_06.jpg", + "4001241_07.jpg", + "4001241_08.jpg", + "4001241_09.jpg", + "4001241_10.jpg", + "4001241_11.jpg", + "4001241_12.jpg", + "4001241_13.jpg", + "4001241_14.jpg", + "4001241_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"bedroom, large windows looking out the city in the future, cinematic, 4k \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001704", + "images": [ + "7001704_00.jpg", + "7001704_01.jpg", + "7001704_02.jpg", + "7001704_03.jpg", + "7001704_04.jpg", + "7001704_05.jpg", + "7001704_06.jpg", + "7001704_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"body camera video of a police officer who comes face to face with a creepy creature, fisheye, realistic, grain. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 2\n" + } + ] + }, + { + "id": "6000545", + "images": [ + "6000545_00.jpg", + "6000545_01.jpg", + "6000545_02.jpg", + "6000545_03.jpg", + "6000545_04.jpg", + "6000545_05.jpg", + "6000545_06.jpg", + "6000545_07.jpg", + "6000545_08.jpg", + "6000545_09.jpg", + "6000545_10.jpg", + "6000545_11.jpg", + "6000545_12.jpg", + "6000545_13.jpg", + "6000545_14.jpg", + "6000545_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Before my bed a pool of night. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "0000234", + "images": [ + "0000234_00.jpg", + "0000234_01.jpg", + "0000234_02.jpg", + "0000234_03.jpg", + "0000234_04.jpg", + "0000234_05.jpg", + "0000234_06.jpg", + "0000234_07.jpg", + "0000234_08.jpg", + "0000234_09.jpg", + "0000234_10.jpg", + "0000234_11.jpg", + "0000234_12.jpg", + "0000234_13.jpg", + "0000234_14.jpg", + "0000234_15.jpg", + "0000234_16.jpg", + "0000234_17.jpg", + "0000234_18.jpg", + "0000234_19.jpg", + "0000234_20.jpg", + "0000234_21.jpg", + "0000234_22.jpg", + "0000234_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"police officers dancing in the street \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000567", + "images": [ + "3000567_00.jpg", + "3000567_01.jpg", + "3000567_02.jpg", + "3000567_03.jpg", + "3000567_04.jpg", + "3000567_05.jpg", + "3000567_06.jpg", + "3000567_07.jpg", + "3000567_08.jpg", + "3000567_09.jpg", + "3000567_10.jpg", + "3000567_11.jpg", + "3000567_12.jpg", + "3000567_13.jpg", + "3000567_14.jpg", + "3000567_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"mermaids in the Soviet Union with scientist \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "4001137", + "images": [ + "4001137_00.jpg", + "4001137_01.jpg", + "4001137_02.jpg", + "4001137_03.jpg", + "4001137_04.jpg", + "4001137_05.jpg", + "4001137_06.jpg", + "4001137_07.jpg", + "4001137_08.jpg", + "4001137_09.jpg", + "4001137_10.jpg", + "4001137_11.jpg", + "4001137_12.jpg", + "4001137_13.jpg", + "4001137_14.jpg", + "4001137_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a snowy street in victorian times at christmas,a coffee shop with a horse and cart outside,its snowy and christmas lights are flashing in windows of homes. The street is cobbled and street lamps are glowing. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002753", + "images": [ + "2002753_00.jpg", + "2002753_01.jpg", + "2002753_02.jpg", + "2002753_03.jpg", + "2002753_04.jpg", + "2002753_05.jpg", + "2002753_06.jpg", + "2002753_07.jpg", + "2002753_08.jpg", + "2002753_09.jpg", + "2002753_10.jpg", + "2002753_11.jpg", + "2002753_12.jpg", + "2002753_13.jpg", + "2002753_14.jpg", + "2002753_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A beautiful woman stands on the top of Victoria Peak in Hong Kong, looking at the night view. There is no one around, and she opens her arms to embrace. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002765", + "images": [ + "2002765_00.jpg", + "2002765_01.jpg", + "2002765_02.jpg", + "2002765_03.jpg", + "2002765_04.jpg", + "2002765_05.jpg", + "2002765_06.jpg", + "2002765_07.jpg", + "2002765_08.jpg", + "2002765_09.jpg", + "2002765_10.jpg", + "2002765_11.jpg", + "2002765_12.jpg", + "2002765_13.jpg", + "2002765_14.jpg", + "2002765_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Robot petting a cat on the background of a full moon \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1000476", + "images": [ + "1000476_00.jpg", + "1000476_01.jpg", + "1000476_02.jpg", + "1000476_03.jpg", + "1000476_04.jpg", + "1000476_05.jpg", + "1000476_06.jpg", + "1000476_07.jpg", + "1000476_08.jpg", + "1000476_09.jpg", + "1000476_10.jpg", + "1000476_11.jpg", + "1000476_12.jpg", + "1000476_13.jpg", + "1000476_14.jpg", + "1000476_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Darth Vader in Star Wars standing on a transparent wet ground watching a spaceship fight shooting green rays on a gray sky high quality video \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0001580", + "images": [ + "0001580_00.jpg", + "0001580_01.jpg", + "0001580_02.jpg", + "0001580_03.jpg", + "0001580_04.jpg", + "0001580_05.jpg", + "0001580_06.jpg", + "0001580_07.jpg", + "0001580_08.jpg", + "0001580_09.jpg", + "0001580_10.jpg", + "0001580_11.jpg", + "0001580_12.jpg", + "0001580_13.jpg", + "0001580_14.jpg", + "0001580_15.jpg", + "0001580_16.jpg", + "0001580_17.jpg", + "0001580_18.jpg", + "0001580_19.jpg", + "0001580_20.jpg", + "0001580_21.jpg", + "0001580_22.jpg", + "0001580_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"one night a boy look moon \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001553", + "images": [ + "1001553_00.jpg", + "1001553_01.jpg", + "1001553_02.jpg", + "1001553_03.jpg", + "1001553_04.jpg", + "1001553_05.jpg", + "1001553_06.jpg", + "1001553_07.jpg", + "1001553_08.jpg", + "1001553_09.jpg", + "1001553_10.jpg", + "1001553_11.jpg", + "1001553_12.jpg", + "1001553_13.jpg", + "1001553_14.jpg", + "1001553_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"he camera pans over a rugged landscape, reminiscent of a jungle or desert terrain, with ancient ruins barely visible in the distance. A dramatic musical score sets the tone for adventure. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3001586", + "images": [ + "3001586_00.jpg", + "3001586_01.jpg", + "3001586_02.jpg", + "3001586_03.jpg", + "3001586_04.jpg", + "3001586_05.jpg", + "3001586_06.jpg", + "3001586_07.jpg", + "3001586_08.jpg", + "3001586_09.jpg", + "3001586_10.jpg", + "3001586_11.jpg", + "3001586_12.jpg", + "3001586_13.jpg", + "3001586_14.jpg", + "3001586_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"dusty street, dusty buildings, burning cars, car crash, dark red sky, eclipse \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "0000430", + "images": [ + "0000430_00.jpg", + "0000430_01.jpg", + "0000430_02.jpg", + "0000430_03.jpg", + "0000430_04.jpg", + "0000430_05.jpg", + "0000430_06.jpg", + "0000430_07.jpg", + "0000430_08.jpg", + "0000430_09.jpg", + "0000430_10.jpg", + "0000430_11.jpg", + "0000430_12.jpg", + "0000430_13.jpg", + "0000430_14.jpg", + "0000430_15.jpg", + "0000430_16.jpg", + "0000430_17.jpg", + "0000430_18.jpg", + "0000430_19.jpg", + "0000430_20.jpg", + "0000430_21.jpg", + "0000430_22.jpg", + "0000430_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Drone approaching an abandoned building from the inside \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0002944", + "images": [ + "0002944_00.jpg", + "0002944_01.jpg", + "0002944_02.jpg", + "0002944_03.jpg", + "0002944_04.jpg", + "0002944_05.jpg", + "0002944_06.jpg", + "0002944_07.jpg", + "0002944_08.jpg", + "0002944_09.jpg", + "0002944_10.jpg", + "0002944_11.jpg", + "0002944_12.jpg", + "0002944_13.jpg", + "0002944_14.jpg", + "0002944_15.jpg", + "0002944_16.jpg", + "0002944_17.jpg", + "0002944_18.jpg", + "0002944_19.jpg", + "0002944_20.jpg", + "0002944_21.jpg", + "0002944_22.jpg", + "0002944_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Take care of the Earth and take care of you \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001399", + "images": [ + "1001399_00.jpg", + "1001399_01.jpg", + "1001399_02.jpg", + "1001399_03.jpg", + "1001399_04.jpg", + "1001399_05.jpg", + "1001399_06.jpg", + "1001399_07.jpg", + "1001399_08.jpg", + "1001399_09.jpg", + "1001399_10.jpg", + "1001399_11.jpg", + "1001399_12.jpg", + "1001399_13.jpg", + "1001399_14.jpg", + "1001399_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Lao Tzu, Tao Te Ching, Landscape, Spring and Autumn Period \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000562", + "images": [ + "3000562_00.jpg", + "3000562_01.jpg", + "3000562_02.jpg", + "3000562_03.jpg", + "3000562_04.jpg", + "3000562_05.jpg", + "3000562_06.jpg", + "3000562_07.jpg", + "3000562_08.jpg", + "3000562_09.jpg", + "3000562_10.jpg", + "3000562_11.jpg", + "3000562_12.jpg", + "3000562_13.jpg", + "3000562_14.jpg", + "3000562_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Move the lens forward, Coffee Shop, Cyberpunk \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002540", + "images": [ + "2002540_00.jpg", + "2002540_01.jpg", + "2002540_02.jpg", + "2002540_03.jpg", + "2002540_04.jpg", + "2002540_05.jpg", + "2002540_06.jpg", + "2002540_07.jpg", + "2002540_08.jpg", + "2002540_09.jpg", + "2002540_10.jpg", + "2002540_11.jpg", + "2002540_12.jpg", + "2002540_13.jpg", + "2002540_14.jpg", + "2002540_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"30 seconds video of sardinian girls having party \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0002738", + "images": [ + "0002738_00.jpg", + "0002738_01.jpg", + "0002738_02.jpg", + "0002738_03.jpg", + "0002738_04.jpg", + "0002738_05.jpg", + "0002738_06.jpg", + "0002738_07.jpg", + "0002738_08.jpg", + "0002738_09.jpg", + "0002738_10.jpg", + "0002738_11.jpg", + "0002738_12.jpg", + "0002738_13.jpg", + "0002738_14.jpg", + "0002738_15.jpg", + "0002738_16.jpg", + "0002738_17.jpg", + "0002738_18.jpg", + "0002738_19.jpg", + "0002738_20.jpg", + "0002738_21.jpg", + "0002738_22.jpg", + "0002738_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"girl wearing violet jeans short swinging an axe 4k resolution \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "8000064", + "images": [ + "8000064_00.jpg", + "8000064_01.jpg", + "8000064_02.jpg", + "8000064_03.jpg", + "8000064_04.jpg", + "8000064_05.jpg", + "8000064_06.jpg", + "8000064_07.jpg", + "8000064_08.jpg", + "8000064_09.jpg", + "8000064_10.jpg", + "8000064_11.jpg", + "8000064_12.jpg", + "8000064_13.jpg", + "8000064_14.jpg", + "8000064_15.jpg", + "8000064_16.jpg", + "8000064_17.jpg", + "8000064_18.jpg", + "8000064_19.jpg", + "8000064_20.jpg", + "8000064_21.jpg", + "8000064_22.jpg", + "8000064_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"zoom in and move in Space ship \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000633", + "images": [ + "3000633_00.jpg", + "3000633_01.jpg", + "3000633_02.jpg", + "3000633_03.jpg", + "3000633_04.jpg", + "3000633_05.jpg", + "3000633_06.jpg", + "3000633_07.jpg", + "3000633_08.jpg", + "3000633_09.jpg", + "3000633_10.jpg", + "3000633_11.jpg", + "3000633_12.jpg", + "3000633_13.jpg", + "3000633_14.jpg", + "3000633_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"girl sleeping in a scary house black and white video \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "4000208", + "images": [ + "4000208_00.jpg", + "4000208_01.jpg", + "4000208_02.jpg", + "4000208_03.jpg", + "4000208_04.jpg", + "4000208_05.jpg", + "4000208_06.jpg", + "4000208_07.jpg", + "4000208_08.jpg", + "4000208_09.jpg", + "4000208_10.jpg", + "4000208_11.jpg", + "4000208_12.jpg", + "4000208_13.jpg", + "4000208_14.jpg", + "4000208_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a technicolor skull with attached spinal column is laying on a bed of plastic moving is lower jaw as if speaking 4k \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0002055", + "images": [ + "0002055_00.jpg", + "0002055_01.jpg", + "0002055_02.jpg", + "0002055_03.jpg", + "0002055_04.jpg", + "0002055_05.jpg", + "0002055_06.jpg", + "0002055_07.jpg", + "0002055_08.jpg", + "0002055_09.jpg", + "0002055_10.jpg", + "0002055_11.jpg", + "0002055_12.jpg", + "0002055_13.jpg", + "0002055_14.jpg", + "0002055_15.jpg", + "0002055_16.jpg", + "0002055_17.jpg", + "0002055_18.jpg", + "0002055_19.jpg", + "0002055_20.jpg", + "0002055_21.jpg", + "0002055_22.jpg", + "0002055_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A cozy living room with Doraemon and Nobita talking \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002186", + "images": [ + "3002186_00.jpg", + "3002186_01.jpg", + "3002186_02.jpg", + "3002186_03.jpg", + "3002186_04.jpg", + "3002186_05.jpg", + "3002186_06.jpg", + "3002186_07.jpg", + "3002186_08.jpg", + "3002186_09.jpg", + "3002186_10.jpg", + "3002186_11.jpg", + "3002186_12.jpg", + "3002186_13.jpg", + "3002186_14.jpg", + "3002186_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"gods spirits and forces of nature \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2001324", + "images": [ + "2001324_00.jpg", + "2001324_01.jpg", + "2001324_02.jpg", + "2001324_03.jpg", + "2001324_04.jpg", + "2001324_05.jpg", + "2001324_06.jpg", + "2001324_07.jpg", + "2001324_08.jpg", + "2001324_09.jpg", + "2001324_10.jpg", + "2001324_11.jpg", + "2001324_12.jpg", + "2001324_13.jpg", + "2001324_14.jpg", + "2001324_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Girl holding a flower and smiling, wind blowing around \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000412", + "images": [ + "5000412_00.jpg", + "5000412_01.jpg", + "5000412_02.jpg", + "5000412_03.jpg", + "5000412_04.jpg", + "5000412_05.jpg", + "5000412_06.jpg", + "5000412_07.jpg", + "5000412_08.jpg", + "5000412_09.jpg", + "5000412_10.jpg", + "5000412_11.jpg", + "5000412_12.jpg", + "5000412_13.jpg", + "5000412_14.jpg", + "5000412_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"1 mom, 1 girl, play, park, photorealistic, 4k, UHD, film, high detailed, Best quality \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001911", + "images": [ + "6001911_00.jpg", + "6001911_01.jpg", + "6001911_02.jpg", + "6001911_03.jpg", + "6001911_04.jpg", + "6001911_05.jpg", + "6001911_06.jpg", + "6001911_07.jpg", + "6001911_08.jpg", + "6001911_09.jpg", + "6001911_10.jpg", + "6001911_11.jpg", + "6001911_12.jpg", + "6001911_13.jpg", + "6001911_14.jpg", + "6001911_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"dancing sushi on the moon doing the floss \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002561", + "images": [ + "3002561_00.jpg", + "3002561_01.jpg", + "3002561_02.jpg", + "3002561_03.jpg", + "3002561_04.jpg", + "3002561_05.jpg", + "3002561_06.jpg", + "3002561_07.jpg", + "3002561_08.jpg", + "3002561_09.jpg", + "3002561_10.jpg", + "3002561_11.jpg", + "3002561_12.jpg", + "3002561_13.jpg", + "3002561_14.jpg", + "3002561_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The man is using facebook and pulling the mobile display \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000288", + "images": [ + "6000288_00.jpg", + "6000288_01.jpg", + "6000288_02.jpg", + "6000288_03.jpg", + "6000288_04.jpg", + "6000288_05.jpg", + "6000288_06.jpg", + "6000288_07.jpg", + "6000288_08.jpg", + "6000288_09.jpg", + "6000288_10.jpg", + "6000288_11.jpg", + "6000288_12.jpg", + "6000288_13.jpg", + "6000288_14.jpg", + "6000288_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a cartoon polar bear with three heads juggles penguins \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3001283", + "images": [ + "3001283_00.jpg", + "3001283_01.jpg", + "3001283_02.jpg", + "3001283_03.jpg", + "3001283_04.jpg", + "3001283_05.jpg", + "3001283_06.jpg", + "3001283_07.jpg", + "3001283_08.jpg", + "3001283_09.jpg", + "3001283_10.jpg", + "3001283_11.jpg", + "3001283_12.jpg", + "3001283_13.jpg", + "3001283_14.jpg", + "3001283_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a friendly talking butterfly with blue and green colors \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0002578", + "images": [ + "0002578_00.jpg", + "0002578_01.jpg", + "0002578_02.jpg", + "0002578_03.jpg", + "0002578_04.jpg", + "0002578_05.jpg", + "0002578_06.jpg", + "0002578_07.jpg", + "0002578_08.jpg", + "0002578_09.jpg", + "0002578_10.jpg", + "0002578_11.jpg", + "0002578_12.jpg", + "0002578_13.jpg", + "0002578_14.jpg", + "0002578_15.jpg", + "0002578_16.jpg", + "0002578_17.jpg", + "0002578_18.jpg", + "0002578_19.jpg", + "0002578_20.jpg", + "0002578_21.jpg", + "0002578_22.jpg", + "0002578_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"55 yo man walking in the park, early morning \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000159", + "images": [ + "4000159_00.jpg", + "4000159_01.jpg", + "4000159_02.jpg", + "4000159_03.jpg", + "4000159_04.jpg", + "4000159_05.jpg", + "4000159_06.jpg", + "4000159_07.jpg", + "4000159_08.jpg", + "4000159_09.jpg", + "4000159_10.jpg", + "4000159_11.jpg", + "4000159_12.jpg", + "4000159_13.jpg", + "4000159_14.jpg", + "4000159_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a couple doing coition on a beach \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "1000583", + "images": [ + "1000583_00.jpg", + "1000583_01.jpg", + "1000583_02.jpg", + "1000583_03.jpg", + "1000583_04.jpg", + "1000583_05.jpg", + "1000583_06.jpg", + "1000583_07.jpg", + "1000583_08.jpg", + "1000583_09.jpg", + "1000583_10.jpg", + "1000583_11.jpg", + "1000583_12.jpg", + "1000583_13.jpg", + "1000583_14.jpg", + "1000583_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A beautiful girl,is sleeping on bed, bikini, memory cotton pillow \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002330", + "images": [ + "2002330_00.jpg", + "2002330_01.jpg", + "2002330_02.jpg", + "2002330_03.jpg", + "2002330_04.jpg", + "2002330_05.jpg", + "2002330_06.jpg", + "2002330_07.jpg", + "2002330_08.jpg", + "2002330_09.jpg", + "2002330_10.jpg", + "2002330_11.jpg", + "2002330_12.jpg", + "2002330_13.jpg", + "2002330_14.jpg", + "2002330_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"login inside a paper plane that opens up \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000037", + "images": [ + "5000037_00.jpg", + "5000037_01.jpg", + "5000037_02.jpg", + "5000037_03.jpg", + "5000037_04.jpg", + "5000037_05.jpg", + "5000037_06.jpg", + "5000037_07.jpg", + "5000037_08.jpg", + "5000037_09.jpg", + "5000037_10.jpg", + "5000037_11.jpg", + "5000037_12.jpg", + "5000037_13.jpg", + "5000037_14.jpg", + "5000037_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"an aerial view of an old wooden house found in a far village, a long alley of trees leading to the house \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000107", + "images": [ + "7000107_00.jpg", + "7000107_01.jpg", + "7000107_02.jpg", + "7000107_03.jpg", + "7000107_04.jpg", + "7000107_05.jpg", + "7000107_06.jpg", + "7000107_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"moses walking to mountain talking to god, hyper realistic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000318", + "images": [ + "3000318_00.jpg", + "3000318_01.jpg", + "3000318_02.jpg", + "3000318_03.jpg", + "3000318_04.jpg", + "3000318_05.jpg", + "3000318_06.jpg", + "3000318_07.jpg", + "3000318_08.jpg", + "3000318_09.jpg", + "3000318_10.jpg", + "3000318_11.jpg", + "3000318_12.jpg", + "3000318_13.jpg", + "3000318_14.jpg", + "3000318_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"3D lion in jungle with golden hair \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002984", + "images": [ + "3002984_00.jpg", + "3002984_01.jpg", + "3002984_02.jpg", + "3002984_03.jpg", + "3002984_04.jpg", + "3002984_05.jpg", + "3002984_06.jpg", + "3002984_07.jpg", + "3002984_08.jpg", + "3002984_09.jpg", + "3002984_10.jpg", + "3002984_11.jpg", + "3002984_12.jpg", + "3002984_13.jpg", + "3002984_14.jpg", + "3002984_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"an anime character dressed like an journalist \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "6000316", + "images": [ + "6000316_00.jpg", + "6000316_01.jpg", + "6000316_02.jpg", + "6000316_03.jpg", + "6000316_04.jpg", + "6000316_05.jpg", + "6000316_06.jpg", + "6000316_07.jpg", + "6000316_08.jpg", + "6000316_09.jpg", + "6000316_10.jpg", + "6000316_11.jpg", + "6000316_12.jpg", + "6000316_13.jpg", + "6000316_14.jpg", + "6000316_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"This is a horseback archery performance. In a sandy indoor performance venue that is 30 meters long and 60 meters wide, a person is riding a horse and shooting at a grass target while being carried on the horseback. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002547", + "images": [ + "2002547_00.jpg", + "2002547_01.jpg", + "2002547_02.jpg", + "2002547_03.jpg", + "2002547_04.jpg", + "2002547_05.jpg", + "2002547_06.jpg", + "2002547_07.jpg", + "2002547_08.jpg", + "2002547_09.jpg", + "2002547_10.jpg", + "2002547_11.jpg", + "2002547_12.jpg", + "2002547_13.jpg", + "2002547_14.jpg", + "2002547_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A Gryphon and a Dragon fighting in the spring season with a little snowfall \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0001732", + "images": [ + "0001732_00.jpg", + "0001732_01.jpg", + "0001732_02.jpg", + "0001732_03.jpg", + "0001732_04.jpg", + "0001732_05.jpg", + "0001732_06.jpg", + "0001732_07.jpg", + "0001732_08.jpg", + "0001732_09.jpg", + "0001732_10.jpg", + "0001732_11.jpg", + "0001732_12.jpg", + "0001732_13.jpg", + "0001732_14.jpg", + "0001732_15.jpg", + "0001732_16.jpg", + "0001732_17.jpg", + "0001732_18.jpg", + "0001732_19.jpg", + "0001732_20.jpg", + "0001732_21.jpg", + "0001732_22.jpg", + "0001732_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"drone footage of film noir city at night, intense light and shadow, neon, moonlight\",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000831", + "images": [ + "5000831_00.jpg", + "5000831_01.jpg", + "5000831_02.jpg", + "5000831_03.jpg", + "5000831_04.jpg", + "5000831_05.jpg", + "5000831_06.jpg", + "5000831_07.jpg", + "5000831_08.jpg", + "5000831_09.jpg", + "5000831_10.jpg", + "5000831_11.jpg", + "5000831_12.jpg", + "5000831_13.jpg", + "5000831_14.jpg", + "5000831_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Picture a pit in the jungle, with a frightened baby elephant inside, struggling to climb out, its eyes pleading for help. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002025", + "images": [ + "2002025_00.jpg", + "2002025_01.jpg", + "2002025_02.jpg", + "2002025_03.jpg", + "2002025_04.jpg", + "2002025_05.jpg", + "2002025_06.jpg", + "2002025_07.jpg", + "2002025_08.jpg", + "2002025_09.jpg", + "2002025_10.jpg", + "2002025_11.jpg", + "2002025_12.jpg", + "2002025_13.jpg", + "2002025_14.jpg", + "2002025_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Ayan shielded his eyes, and when the light subsided, he found himself surrounded by mystical creatures and magical beings. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "1000656", + "images": [ + "1000656_00.jpg", + "1000656_01.jpg", + "1000656_02.jpg", + "1000656_03.jpg", + "1000656_04.jpg", + "1000656_05.jpg", + "1000656_06.jpg", + "1000656_07.jpg", + "1000656_08.jpg", + "1000656_09.jpg", + "1000656_10.jpg", + "1000656_11.jpg", + "1000656_12.jpg", + "1000656_13.jpg", + "1000656_14.jpg", + "1000656_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"dusiness center on blue sky Message: custdevinterview.ru (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002322", + "images": [ + "2002322_00.jpg", + "2002322_01.jpg", + "2002322_02.jpg", + "2002322_03.jpg", + "2002322_04.jpg", + "2002322_05.jpg", + "2002322_06.jpg", + "2002322_07.jpg", + "2002322_08.jpg", + "2002322_09.jpg", + "2002322_10.jpg", + "2002322_11.jpg", + "2002322_12.jpg", + "2002322_13.jpg", + "2002322_14.jpg", + "2002322_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"realistic movie scene depicting a horror masked ball \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001000", + "images": [ + "7001000_00.jpg", + "7001000_01.jpg", + "7001000_02.jpg", + "7001000_03.jpg", + "7001000_04.jpg", + "7001000_05.jpg", + "7001000_06.jpg", + "7001000_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Create a realistic image of the the pyramids of Meroe standing tall against the backdrop of the setting sun. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001357", + "images": [ + "1001357_00.jpg", + "1001357_01.jpg", + "1001357_02.jpg", + "1001357_03.jpg", + "1001357_04.jpg", + "1001357_05.jpg", + "1001357_06.jpg", + "1001357_07.jpg", + "1001357_08.jpg", + "1001357_09.jpg", + "1001357_10.jpg", + "1001357_11.jpg", + "1001357_12.jpg", + "1001357_13.jpg", + "1001357_14.jpg", + "1001357_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"plan view of round vibrant led city \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000269", + "images": [ + "6000269_00.jpg", + "6000269_01.jpg", + "6000269_02.jpg", + "6000269_03.jpg", + "6000269_04.jpg", + "6000269_05.jpg", + "6000269_06.jpg", + "6000269_07.jpg", + "6000269_08.jpg", + "6000269_09.jpg", + "6000269_10.jpg", + "6000269_11.jpg", + "6000269_12.jpg", + "6000269_13.jpg", + "6000269_14.jpg", + "6000269_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"an helenistic sculpture with digital glitchs errors spining slowly side to side in a vaporwave style. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002782", + "images": [ + "2002782_00.jpg", + "2002782_01.jpg", + "2002782_02.jpg", + "2002782_03.jpg", + "2002782_04.jpg", + "2002782_05.jpg", + "2002782_06.jpg", + "2002782_07.jpg", + "2002782_08.jpg", + "2002782_09.jpg", + "2002782_10.jpg", + "2002782_11.jpg", + "2002782_12.jpg", + "2002782_13.jpg", + "2002782_14.jpg", + "2002782_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"create an animated scene with a young girl missing her foot while walking in a living root bridge \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "1000652", + "images": [ + "1000652_00.jpg", + "1000652_01.jpg", + "1000652_02.jpg", + "1000652_03.jpg", + "1000652_04.jpg", + "1000652_05.jpg", + "1000652_06.jpg", + "1000652_07.jpg", + "1000652_08.jpg", + "1000652_09.jpg", + "1000652_10.jpg", + "1000652_11.jpg", + "1000652_12.jpg", + "1000652_13.jpg", + "1000652_14.jpg", + "1000652_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a man opening the wooded door \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000393", + "images": [ + "6000393_00.jpg", + "6000393_01.jpg", + "6000393_02.jpg", + "6000393_03.jpg", + "6000393_04.jpg", + "6000393_05.jpg", + "6000393_06.jpg", + "6000393_07.jpg", + "6000393_08.jpg", + "6000393_09.jpg", + "6000393_10.jpg", + "6000393_11.jpg", + "6000393_12.jpg", + "6000393_13.jpg", + "6000393_14.jpg", + "6000393_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a young man with solid personality walking in tothe thick forest with back to the camera from down to up ,hd.4k resolution \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1000033", + "images": [ + "1000033_00.jpg", + "1000033_01.jpg", + "1000033_02.jpg", + "1000033_03.jpg", + "1000033_04.jpg", + "1000033_05.jpg", + "1000033_06.jpg", + "1000033_07.jpg", + "1000033_08.jpg", + "1000033_09.jpg", + "1000033_10.jpg", + "1000033_11.jpg", + "1000033_12.jpg", + "1000033_13.jpg", + "1000033_14.jpg", + "1000033_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"god and jesus are standing in the street of ancient city. street view. realistic immage \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0002323", + "images": [ + "0002323_00.jpg", + "0002323_01.jpg", + "0002323_02.jpg", + "0002323_03.jpg", + "0002323_04.jpg", + "0002323_05.jpg", + "0002323_06.jpg", + "0002323_07.jpg", + "0002323_08.jpg", + "0002323_09.jpg", + "0002323_10.jpg", + "0002323_11.jpg", + "0002323_12.jpg", + "0002323_13.jpg", + "0002323_14.jpg", + "0002323_15.jpg", + "0002323_16.jpg", + "0002323_17.jpg", + "0002323_18.jpg", + "0002323_19.jpg", + "0002323_20.jpg", + "0002323_21.jpg", + "0002323_22.jpg", + "0002323_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"two women fighting in a street market \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7000018", + "images": [ + "7000018_00.jpg", + "7000018_01.jpg", + "7000018_02.jpg", + "7000018_03.jpg", + "7000018_04.jpg", + "7000018_05.jpg", + "7000018_06.jpg", + "7000018_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Roman soldier in Amazon rainforest. Hyperrealism. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7000117", + "images": [ + "7000117_00.jpg", + "7000117_01.jpg", + "7000117_02.jpg", + "7000117_03.jpg", + "7000117_04.jpg", + "7000117_05.jpg", + "7000117_06.jpg", + "7000117_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a rainy scene with a cactus growing \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001824", + "images": [ + "7001824_00.jpg", + "7001824_01.jpg", + "7001824_02.jpg", + "7001824_03.jpg", + "7001824_04.jpg", + "7001824_05.jpg", + "7001824_06.jpg", + "7001824_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a to z create animated video for childrens \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002237", + "images": [ + "3002237_00.jpg", + "3002237_01.jpg", + "3002237_02.jpg", + "3002237_03.jpg", + "3002237_04.jpg", + "3002237_05.jpg", + "3002237_06.jpg", + "3002237_07.jpg", + "3002237_08.jpg", + "3002237_09.jpg", + "3002237_10.jpg", + "3002237_11.jpg", + "3002237_12.jpg", + "3002237_13.jpg", + "3002237_14.jpg", + "3002237_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a closeup detailed wild life potrait of baby elephant, 4k, cinematic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0002910", + "images": [ + "0002910_00.jpg", + "0002910_01.jpg", + "0002910_02.jpg", + "0002910_03.jpg", + "0002910_04.jpg", + "0002910_05.jpg", + "0002910_06.jpg", + "0002910_07.jpg", + "0002910_08.jpg", + "0002910_09.jpg", + "0002910_10.jpg", + "0002910_11.jpg", + "0002910_12.jpg", + "0002910_13.jpg", + "0002910_14.jpg", + "0002910_15.jpg", + "0002910_16.jpg", + "0002910_17.jpg", + "0002910_18.jpg", + "0002910_19.jpg", + "0002910_20.jpg", + "0002910_21.jpg", + "0002910_22.jpg", + "0002910_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"make me a sky of rain and thunder \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001078", + "images": [ + "7001078_00.jpg", + "7001078_01.jpg", + "7001078_02.jpg", + "7001078_03.jpg", + "7001078_04.jpg", + "7001078_05.jpg", + "7001078_06.jpg", + "7001078_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The old man is swinging in the park, 4k, sharp, Arri Alexa, depth of field, highly detailed \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002539", + "images": [ + "3002539_00.jpg", + "3002539_01.jpg", + "3002539_02.jpg", + "3002539_03.jpg", + "3002539_04.jpg", + "3002539_05.jpg", + "3002539_06.jpg", + "3002539_07.jpg", + "3002539_08.jpg", + "3002539_09.jpg", + "3002539_10.jpg", + "3002539_11.jpg", + "3002539_12.jpg", + "3002539_13.jpg", + "3002539_14.jpg", + "3002539_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Realistic video recorded of a cafe in Paris \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000406", + "images": [ + "6000406_00.jpg", + "6000406_01.jpg", + "6000406_02.jpg", + "6000406_03.jpg", + "6000406_04.jpg", + "6000406_05.jpg", + "6000406_06.jpg", + "6000406_07.jpg", + "6000406_08.jpg", + "6000406_09.jpg", + "6000406_10.jpg", + "6000406_11.jpg", + "6000406_12.jpg", + "6000406_13.jpg", + "6000406_14.jpg", + "6000406_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"In a dense forest, Timmy and Lily, holding a treasure map, stand at the edge of a sparkling river. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000861", + "images": [ + "3000861_00.jpg", + "3000861_01.jpg", + "3000861_02.jpg", + "3000861_03.jpg", + "3000861_04.jpg", + "3000861_05.jpg", + "3000861_06.jpg", + "3000861_07.jpg", + "3000861_08.jpg", + "3000861_09.jpg", + "3000861_10.jpg", + "3000861_11.jpg", + "3000861_12.jpg", + "3000861_13.jpg", + "3000861_14.jpg", + "3000861_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"hand draws a portrait of a beautiful girl with pencils, clear lines, beautiful picture, realism, real portrait, artist, high quality \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "5000677", + "images": [ + "5000677_00.jpg", + "5000677_01.jpg", + "5000677_02.jpg", + "5000677_03.jpg", + "5000677_04.jpg", + "5000677_05.jpg", + "5000677_06.jpg", + "5000677_07.jpg", + "5000677_08.jpg", + "5000677_09.jpg", + "5000677_10.jpg", + "5000677_11.jpg", + "5000677_12.jpg", + "5000677_13.jpg", + "5000677_14.jpg", + "5000677_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"anime style , closeup shot, man and woman handshake , Studio Mappa Style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0002721", + "images": [ + "0002721_00.jpg", + "0002721_01.jpg", + "0002721_02.jpg", + "0002721_03.jpg", + "0002721_04.jpg", + "0002721_05.jpg", + "0002721_06.jpg", + "0002721_07.jpg", + "0002721_08.jpg", + "0002721_09.jpg", + "0002721_10.jpg", + "0002721_11.jpg", + "0002721_12.jpg", + "0002721_13.jpg", + "0002721_14.jpg", + "0002721_15.jpg", + "0002721_16.jpg", + "0002721_17.jpg", + "0002721_18.jpg", + "0002721_19.jpg", + "0002721_20.jpg", + "0002721_21.jpg", + "0002721_22.jpg", + "0002721_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a pregnant woman talking to a handset at the end of a long cord tie with a fetus in her womb \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001510", + "images": [ + "6001510_00.jpg", + "6001510_01.jpg", + "6001510_02.jpg", + "6001510_03.jpg", + "6001510_04.jpg", + "6001510_05.jpg", + "6001510_06.jpg", + "6001510_07.jpg", + "6001510_08.jpg", + "6001510_09.jpg", + "6001510_10.jpg", + "6001510_11.jpg", + "6001510_12.jpg", + "6001510_13.jpg", + "6001510_14.jpg", + "6001510_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a 1920 george melies super8 film about witches cult burning ceremony \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "7001338", + "images": [ + "7001338_00.jpg", + "7001338_01.jpg", + "7001338_02.jpg", + "7001338_03.jpg", + "7001338_04.jpg", + "7001338_05.jpg", + "7001338_06.jpg", + "7001338_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Create the image of an illuminated holy book opening and showing the Hebrew words 8k in cinematic style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1000596", + "images": [ + "1000596_00.jpg", + "1000596_01.jpg", + "1000596_02.jpg", + "1000596_03.jpg", + "1000596_04.jpg", + "1000596_05.jpg", + "1000596_06.jpg", + "1000596_07.jpg", + "1000596_08.jpg", + "1000596_09.jpg", + "1000596_10.jpg", + "1000596_11.jpg", + "1000596_12.jpg", + "1000596_13.jpg", + "1000596_14.jpg", + "1000596_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Thirty second video of a house with a sunset rising behind it \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001604", + "images": [ + "6001604_00.jpg", + "6001604_01.jpg", + "6001604_02.jpg", + "6001604_03.jpg", + "6001604_04.jpg", + "6001604_05.jpg", + "6001604_06.jpg", + "6001604_07.jpg", + "6001604_08.jpg", + "6001604_09.jpg", + "6001604_10.jpg", + "6001604_11.jpg", + "6001604_12.jpg", + "6001604_13.jpg", + "6001604_14.jpg", + "6001604_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Super advertising sign , neon green with the image of neatly and tastefully glowing contours of a sports car at night against the backdrop of a megalopolis \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "1000993", + "images": [ + "1000993_00.jpg", + "1000993_01.jpg", + "1000993_02.jpg", + "1000993_03.jpg", + "1000993_04.jpg", + "1000993_05.jpg", + "1000993_06.jpg", + "1000993_07.jpg", + "1000993_08.jpg", + "1000993_09.jpg", + "1000993_10.jpg", + "1000993_11.jpg", + "1000993_12.jpg", + "1000993_13.jpg", + "1000993_14.jpg", + "1000993_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a taking sushi boat loaded with sushi \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3001062", + "images": [ + "3001062_00.jpg", + "3001062_01.jpg", + "3001062_02.jpg", + "3001062_03.jpg", + "3001062_04.jpg", + "3001062_05.jpg", + "3001062_06.jpg", + "3001062_07.jpg", + "3001062_08.jpg", + "3001062_09.jpg", + "3001062_10.jpg", + "3001062_11.jpg", + "3001062_12.jpg", + "3001062_13.jpg", + "3001062_14.jpg", + "3001062_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The villagers, having sensed her presence, rallied together, armed with holy water and crosses, their faith serving as a shield against evil. 4K high definition 2:3 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "1000705", + "images": [ + "1000705_00.jpg", + "1000705_01.jpg", + "1000705_02.jpg", + "1000705_03.jpg", + "1000705_04.jpg", + "1000705_05.jpg", + "1000705_06.jpg", + "1000705_07.jpg", + "1000705_08.jpg", + "1000705_09.jpg", + "1000705_10.jpg", + "1000705_11.jpg", + "1000705_12.jpg", + "1000705_13.jpg", + "1000705_14.jpg", + "1000705_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"face of a young boy under water looking up at the viewer \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002504", + "images": [ + "3002504_00.jpg", + "3002504_01.jpg", + "3002504_02.jpg", + "3002504_03.jpg", + "3002504_04.jpg", + "3002504_05.jpg", + "3002504_06.jpg", + "3002504_07.jpg", + "3002504_08.jpg", + "3002504_09.jpg", + "3002504_10.jpg", + "3002504_11.jpg", + "3002504_12.jpg", + "3002504_13.jpg", + "3002504_14.jpg", + "3002504_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The backrooms. He cautiously explores, scanning the surroundings. Suddenly, the flashlight flickers and dies. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "4000524", + "images": [ + "4000524_00.jpg", + "4000524_01.jpg", + "4000524_02.jpg", + "4000524_03.jpg", + "4000524_04.jpg", + "4000524_05.jpg", + "4000524_06.jpg", + "4000524_07.jpg", + "4000524_08.jpg", + "4000524_09.jpg", + "4000524_10.jpg", + "4000524_11.jpg", + "4000524_12.jpg", + "4000524_13.jpg", + "4000524_14.jpg", + "4000524_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a beautiful woman wearing a futuristic purple outfit catwalking at new york fashion week \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002267", + "images": [ + "1002267_00.jpg", + "1002267_01.jpg", + "1002267_02.jpg", + "1002267_03.jpg", + "1002267_04.jpg", + "1002267_05.jpg", + "1002267_06.jpg", + "1002267_07.jpg", + "1002267_08.jpg", + "1002267_09.jpg", + "1002267_10.jpg", + "1002267_11.jpg", + "1002267_12.jpg", + "1002267_13.jpg", + "1002267_14.jpg", + "1002267_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Driving inside a car with a view of London from the window \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000308", + "images": [ + "7000308_00.jpg", + "7000308_01.jpg", + "7000308_02.jpg", + "7000308_03.jpg", + "7000308_04.jpg", + "7000308_05.jpg", + "7000308_06.jpg", + "7000308_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a cartoon mushroom being picked by a wizard \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "8000579", + "images": [ + "8000579_00.jpg", + "8000579_01.jpg", + "8000579_02.jpg", + "8000579_03.jpg", + "8000579_04.jpg", + "8000579_05.jpg", + "8000579_06.jpg", + "8000579_07.jpg", + "8000579_08.jpg", + "8000579_09.jpg", + "8000579_10.jpg", + "8000579_11.jpg", + "8000579_12.jpg", + "8000579_13.jpg", + "8000579_14.jpg", + "8000579_15.jpg", + "8000579_16.jpg", + "8000579_17.jpg", + "8000579_18.jpg", + "8000579_19.jpg", + "8000579_20.jpg", + "8000579_21.jpg", + "8000579_22.jpg", + "8000579_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a huge gold bar in the spot light in the dark \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "2000098", + "images": [ + "2000098_00.jpg", + "2000098_01.jpg", + "2000098_02.jpg", + "2000098_03.jpg", + "2000098_04.jpg", + "2000098_05.jpg", + "2000098_06.jpg", + "2000098_07.jpg", + "2000098_08.jpg", + "2000098_09.jpg", + "2000098_10.jpg", + "2000098_11.jpg", + "2000098_12.jpg", + "2000098_13.jpg", + "2000098_14.jpg", + "2000098_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"pepe sailing on a boat, sea is moving, clouds are moving, sails are moving, pepe is smiling \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002333", + "images": [ + "3002333_00.jpg", + "3002333_01.jpg", + "3002333_02.jpg", + "3002333_03.jpg", + "3002333_04.jpg", + "3002333_05.jpg", + "3002333_06.jpg", + "3002333_07.jpg", + "3002333_08.jpg", + "3002333_09.jpg", + "3002333_10.jpg", + "3002333_11.jpg", + "3002333_12.jpg", + "3002333_13.jpg", + "3002333_14.jpg", + "3002333_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Moments of peril and bravery are depicted as a young hobbit, Frodo and his companions face challenges that test their courage and loyalty. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "0001608", + "images": [ + "0001608_00.jpg", + "0001608_01.jpg", + "0001608_02.jpg", + "0001608_03.jpg", + "0001608_04.jpg", + "0001608_05.jpg", + "0001608_06.jpg", + "0001608_07.jpg", + "0001608_08.jpg", + "0001608_09.jpg", + "0001608_10.jpg", + "0001608_11.jpg", + "0001608_12.jpg", + "0001608_13.jpg", + "0001608_14.jpg", + "0001608_15.jpg", + "0001608_16.jpg", + "0001608_17.jpg", + "0001608_18.jpg", + "0001608_19.jpg", + "0001608_20.jpg", + "0001608_21.jpg", + "0001608_22.jpg", + "0001608_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Imperial Star Destroyer in space over a stormy planet, cinematic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000795", + "images": [ + "5000795_00.jpg", + "5000795_01.jpg", + "5000795_02.jpg", + "5000795_03.jpg", + "5000795_04.jpg", + "5000795_05.jpg", + "5000795_06.jpg", + "5000795_07.jpg", + "5000795_08.jpg", + "5000795_09.jpg", + "5000795_10.jpg", + "5000795_11.jpg", + "5000795_12.jpg", + "5000795_13.jpg", + "5000795_14.jpg", + "5000795_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A panoramic view of the village at dusk, with the magical lights illuminating the houses, creating a dreamy and otherworldly atmosphere. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000156", + "images": [ + "6000156_00.jpg", + "6000156_01.jpg", + "6000156_02.jpg", + "6000156_03.jpg", + "6000156_04.jpg", + "6000156_05.jpg", + "6000156_06.jpg", + "6000156_07.jpg", + "6000156_08.jpg", + "6000156_09.jpg", + "6000156_10.jpg", + "6000156_11.jpg", + "6000156_12.jpg", + "6000156_13.jpg", + "6000156_14.jpg", + "6000156_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a surreal hyperrealistic dark forest on either side and a waterfall of green matrix code off in the distance. the sky is black with a few stars, nebulae, and green northern lights,4k,cinematic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "8000234", + "images": [ + "8000234_00.jpg", + "8000234_01.jpg", + "8000234_02.jpg", + "8000234_03.jpg", + "8000234_04.jpg", + "8000234_05.jpg", + "8000234_06.jpg", + "8000234_07.jpg", + "8000234_08.jpg", + "8000234_09.jpg", + "8000234_10.jpg", + "8000234_11.jpg", + "8000234_12.jpg", + "8000234_13.jpg", + "8000234_14.jpg", + "8000234_15.jpg", + "8000234_16.jpg", + "8000234_17.jpg", + "8000234_18.jpg", + "8000234_19.jpg", + "8000234_20.jpg", + "8000234_21.jpg", + "8000234_22.jpg", + "8000234_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Slow, melancholic piano tune emphasizing the struggle and discomfort of wearing old shoes. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0000537", + "images": [ + "0000537_00.jpg", + "0000537_01.jpg", + "0000537_02.jpg", + "0000537_03.jpg", + "0000537_04.jpg", + "0000537_05.jpg", + "0000537_06.jpg", + "0000537_07.jpg", + "0000537_08.jpg", + "0000537_09.jpg", + "0000537_10.jpg", + "0000537_11.jpg", + "0000537_12.jpg", + "0000537_13.jpg", + "0000537_14.jpg", + "0000537_15.jpg", + "0000537_16.jpg", + "0000537_17.jpg", + "0000537_18.jpg", + "0000537_19.jpg", + "0000537_20.jpg", + "0000537_21.jpg", + "0000537_22.jpg", + "0000537_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Biblical story, Joseph receiving a colorful tunic from his father, arid land, biblical times, period clothes, ultra realistic, 8k \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002491", + "images": [ + "1002491_00.jpg", + "1002491_01.jpg", + "1002491_02.jpg", + "1002491_03.jpg", + "1002491_04.jpg", + "1002491_05.jpg", + "1002491_06.jpg", + "1002491_07.jpg", + "1002491_08.jpg", + "1002491_09.jpg", + "1002491_10.jpg", + "1002491_11.jpg", + "1002491_12.jpg", + "1002491_13.jpg", + "1002491_14.jpg", + "1002491_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Pope Francis as Neo of the Matrix \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "0001921", + "images": [ + "0001921_00.jpg", + "0001921_01.jpg", + "0001921_02.jpg", + "0001921_03.jpg", + "0001921_04.jpg", + "0001921_05.jpg", + "0001921_06.jpg", + "0001921_07.jpg", + "0001921_08.jpg", + "0001921_09.jpg", + "0001921_10.jpg", + "0001921_11.jpg", + "0001921_12.jpg", + "0001921_13.jpg", + "0001921_14.jpg", + "0001921_15.jpg", + "0001921_16.jpg", + "0001921_17.jpg", + "0001921_18.jpg", + "0001921_19.jpg", + "0001921_20.jpg", + "0001921_21.jpg", + "0001921_22.jpg", + "0001921_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"anime city 4k moving fast sunset \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000081", + "images": [ + "5000081_00.jpg", + "5000081_01.jpg", + "5000081_02.jpg", + "5000081_03.jpg", + "5000081_04.jpg", + "5000081_05.jpg", + "5000081_06.jpg", + "5000081_07.jpg", + "5000081_08.jpg", + "5000081_09.jpg", + "5000081_10.jpg", + "5000081_11.jpg", + "5000081_12.jpg", + "5000081_13.jpg", + "5000081_14.jpg", + "5000081_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"presence of dwaraka nagar under water with lord krishna \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1001413", + "images": [ + "1001413_00.jpg", + "1001413_01.jpg", + "1001413_02.jpg", + "1001413_03.jpg", + "1001413_04.jpg", + "1001413_05.jpg", + "1001413_06.jpg", + "1001413_07.jpg", + "1001413_08.jpg", + "1001413_09.jpg", + "1001413_10.jpg", + "1001413_11.jpg", + "1001413_12.jpg", + "1001413_13.jpg", + "1001413_14.jpg", + "1001413_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"people see food falling on the Sky \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "8000361", + "images": [ + "8000361_00.jpg", + "8000361_01.jpg", + "8000361_02.jpg", + "8000361_03.jpg", + "8000361_04.jpg", + "8000361_05.jpg", + "8000361_06.jpg", + "8000361_07.jpg", + "8000361_08.jpg", + "8000361_09.jpg", + "8000361_10.jpg", + "8000361_11.jpg", + "8000361_12.jpg", + "8000361_13.jpg", + "8000361_14.jpg", + "8000361_15.jpg", + "8000361_16.jpg", + "8000361_17.jpg", + "8000361_18.jpg", + "8000361_19.jpg", + "8000361_20.jpg", + "8000361_21.jpg", + "8000361_22.jpg", + "8000361_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"mermaid on land gasping for air \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000837", + "images": [ + "7000837_00.jpg", + "7000837_01.jpg", + "7000837_02.jpg", + "7000837_03.jpg", + "7000837_04.jpg", + "7000837_05.jpg", + "7000837_06.jpg", + "7000837_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Intense moments of adrenaline surge as the adventurers face numerous obstacles, racing against time to fix the car and escape the perils of the jungle. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 2\n" + } + ] + }, + { + "id": "8000330", + "images": [ + "8000330_00.jpg", + "8000330_01.jpg", + "8000330_02.jpg", + "8000330_03.jpg", + "8000330_04.jpg", + "8000330_05.jpg", + "8000330_06.jpg", + "8000330_07.jpg", + "8000330_08.jpg", + "8000330_09.jpg", + "8000330_10.jpg", + "8000330_11.jpg", + "8000330_12.jpg", + "8000330_13.jpg", + "8000330_14.jpg", + "8000330_15.jpg", + "8000330_16.jpg", + "8000330_17.jpg", + "8000330_18.jpg", + "8000330_19.jpg", + "8000330_20.jpg", + "8000330_21.jpg", + "8000330_22.jpg", + "8000330_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"An upclose cinematic shot of jesus christ walking outside the colosseum \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001148", + "images": [ + "6001148_00.jpg", + "6001148_01.jpg", + "6001148_02.jpg", + "6001148_03.jpg", + "6001148_04.jpg", + "6001148_05.jpg", + "6001148_06.jpg", + "6001148_07.jpg", + "6001148_08.jpg", + "6001148_09.jpg", + "6001148_10.jpg", + "6001148_11.jpg", + "6001148_12.jpg", + "6001148_13.jpg", + "6001148_14.jpg", + "6001148_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"man 50 years old kneeling front \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3001141", + "images": [ + "3001141_00.jpg", + "3001141_01.jpg", + "3001141_02.jpg", + "3001141_03.jpg", + "3001141_04.jpg", + "3001141_05.jpg", + "3001141_06.jpg", + "3001141_07.jpg", + "3001141_08.jpg", + "3001141_09.jpg", + "3001141_10.jpg", + "3001141_11.jpg", + "3001141_12.jpg", + "3001141_13.jpg", + "3001141_14.jpg", + "3001141_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Delhi Police Inspector and Constables atanding near police vehicle \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "8000101", + "images": [ + "8000101_00.jpg", + "8000101_01.jpg", + "8000101_02.jpg", + "8000101_03.jpg", + "8000101_04.jpg", + "8000101_05.jpg", + "8000101_06.jpg", + "8000101_07.jpg", + "8000101_08.jpg", + "8000101_09.jpg", + "8000101_10.jpg", + "8000101_11.jpg", + "8000101_12.jpg", + "8000101_13.jpg", + "8000101_14.jpg", + "8000101_15.jpg", + "8000101_16.jpg", + "8000101_17.jpg", + "8000101_18.jpg", + "8000101_19.jpg", + "8000101_20.jpg", + "8000101_21.jpg", + "8000101_22.jpg", + "8000101_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Strange aliens run through the sky \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001785", + "images": [ + "6001785_00.jpg", + "6001785_01.jpg", + "6001785_02.jpg", + "6001785_03.jpg", + "6001785_04.jpg", + "6001785_05.jpg", + "6001785_06.jpg", + "6001785_07.jpg", + "6001785_08.jpg", + "6001785_09.jpg", + "6001785_10.jpg", + "6001785_11.jpg", + "6001785_12.jpg", + "6001785_13.jpg", + "6001785_14.jpg", + "6001785_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"dancing metal band on stage, exploding guitars \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1002736", + "images": [ + "1002736_00.jpg", + "1002736_01.jpg", + "1002736_02.jpg", + "1002736_03.jpg", + "1002736_04.jpg", + "1002736_05.jpg", + "1002736_06.jpg", + "1002736_07.jpg", + "1002736_08.jpg", + "1002736_09.jpg", + "1002736_10.jpg", + "1002736_11.jpg", + "1002736_12.jpg", + "1002736_13.jpg", + "1002736_14.jpg", + "1002736_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"2 Wookies drinking milk from a clear glass in a modern home kitchen \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "7001343", + "images": [ + "7001343_00.jpg", + "7001343_01.jpg", + "7001343_02.jpg", + "7001343_03.jpg", + "7001343_04.jpg", + "7001343_05.jpg", + "7001343_06.jpg", + "7001343_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a Man take a Ice cream in old Villager 16:9 motion 2 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2000534", + "images": [ + "2000534_00.jpg", + "2000534_01.jpg", + "2000534_02.jpg", + "2000534_03.jpg", + "2000534_04.jpg", + "2000534_05.jpg", + "2000534_06.jpg", + "2000534_07.jpg", + "2000534_08.jpg", + "2000534_09.jpg", + "2000534_10.jpg", + "2000534_11.jpg", + "2000534_12.jpg", + "2000534_13.jpg", + "2000534_14.jpg", + "2000534_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"cinematic videos, hd, 4k, realistic, sharp, \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001711", + "images": [ + "6001711_00.jpg", + "6001711_01.jpg", + "6001711_02.jpg", + "6001711_03.jpg", + "6001711_04.jpg", + "6001711_05.jpg", + "6001711_06.jpg", + "6001711_07.jpg", + "6001711_08.jpg", + "6001711_09.jpg", + "6001711_10.jpg", + "6001711_11.jpg", + "6001711_12.jpg", + "6001711_13.jpg", + "6001711_14.jpg", + "6001711_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"anime man singing song , crowd is cherring motion 4 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "2001027", + "images": [ + "2001027_00.jpg", + "2001027_01.jpg", + "2001027_02.jpg", + "2001027_03.jpg", + "2001027_04.jpg", + "2001027_05.jpg", + "2001027_06.jpg", + "2001027_07.jpg", + "2001027_08.jpg", + "2001027_09.jpg", + "2001027_10.jpg", + "2001027_11.jpg", + "2001027_12.jpg", + "2001027_13.jpg", + "2001027_14.jpg", + "2001027_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"animated video of the fusion of multivesicular bodies with the plasma membrane \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1000001", + "images": [ + "1000001_00.jpg", + "1000001_01.jpg", + "1000001_02.jpg", + "1000001_03.jpg", + "1000001_04.jpg", + "1000001_05.jpg", + "1000001_06.jpg", + "1000001_07.jpg", + "1000001_08.jpg", + "1000001_09.jpg", + "1000001_10.jpg", + "1000001_11.jpg", + "1000001_12.jpg", + "1000001_13.jpg", + "1000001_14.jpg", + "1000001_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"An angel who can\u2019t see her face from the wing in the leaky sky blows into the horn \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3001504", + "images": [ + "3001504_00.jpg", + "3001504_01.jpg", + "3001504_02.jpg", + "3001504_03.jpg", + "3001504_04.jpg", + "3001504_05.jpg", + "3001504_06.jpg", + "3001504_07.jpg", + "3001504_08.jpg", + "3001504_09.jpg", + "3001504_10.jpg", + "3001504_11.jpg", + "3001504_12.jpg", + "3001504_13.jpg", + "3001504_14.jpg", + "3001504_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Dazzling Fanfare: Showcase a brief, vibrant performance where trumpets blast in unison, releasing a dazzling fanfare of music and colorful confetti. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "1002565", + "images": [ + "1002565_00.jpg", + "1002565_01.jpg", + "1002565_02.jpg", + "1002565_03.jpg", + "1002565_04.jpg", + "1002565_05.jpg", + "1002565_06.jpg", + "1002565_07.jpg", + "1002565_08.jpg", + "1002565_09.jpg", + "1002565_10.jpg", + "1002565_11.jpg", + "1002565_12.jpg", + "1002565_13.jpg", + "1002565_14.jpg", + "1002565_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The magical kite led a kids Mia, Ben and leo to a land of candy trees, where they feasted on chocolate leaves and gumdrop fruits. Cinematic, cartoon style. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0000812", + "images": [ + "0000812_00.jpg", + "0000812_01.jpg", + "0000812_02.jpg", + "0000812_03.jpg", + "0000812_04.jpg", + "0000812_05.jpg", + "0000812_06.jpg", + "0000812_07.jpg", + "0000812_08.jpg", + "0000812_09.jpg", + "0000812_10.jpg", + "0000812_11.jpg", + "0000812_12.jpg", + "0000812_13.jpg", + "0000812_14.jpg", + "0000812_15.jpg", + "0000812_16.jpg", + "0000812_17.jpg", + "0000812_18.jpg", + "0000812_19.jpg", + "0000812_20.jpg", + "0000812_21.jpg", + "0000812_22.jpg", + "0000812_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a woman feels tired after a whole day work, she sits down on the chair and then fulls in sleep \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "0001540", + "images": [ + "0001540_00.jpg", + "0001540_01.jpg", + "0001540_02.jpg", + "0001540_03.jpg", + "0001540_04.jpg", + "0001540_05.jpg", + "0001540_06.jpg", + "0001540_07.jpg", + "0001540_08.jpg", + "0001540_09.jpg", + "0001540_10.jpg", + "0001540_11.jpg", + "0001540_12.jpg", + "0001540_13.jpg", + "0001540_14.jpg", + "0001540_15.jpg", + "0001540_16.jpg", + "0001540_17.jpg", + "0001540_18.jpg", + "0001540_19.jpg", + "0001540_20.jpg", + "0001540_21.jpg", + "0001540_22.jpg", + "0001540_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Astronaut sitting in front of the computer and the codes reflected on his helmet. in the dark \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "0001386", + "images": [ + "0001386_00.jpg", + "0001386_01.jpg", + "0001386_02.jpg", + "0001386_03.jpg", + "0001386_04.jpg", + "0001386_05.jpg", + "0001386_06.jpg", + "0001386_07.jpg", + "0001386_08.jpg", + "0001386_09.jpg", + "0001386_10.jpg", + "0001386_11.jpg", + "0001386_12.jpg", + "0001386_13.jpg", + "0001386_14.jpg", + "0001386_15.jpg", + "0001386_16.jpg", + "0001386_17.jpg", + "0001386_18.jpg", + "0001386_19.jpg", + "0001386_20.jpg", + "0001386_21.jpg", + "0001386_22.jpg", + "0001386_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Celebrate Achievements: Acknowledge every small success; it keeps you motivated. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002433", + "images": [ + "2002433_00.jpg", + "2002433_01.jpg", + "2002433_02.jpg", + "2002433_03.jpg", + "2002433_04.jpg", + "2002433_05.jpg", + "2002433_06.jpg", + "2002433_07.jpg", + "2002433_08.jpg", + "2002433_09.jpg", + "2002433_10.jpg", + "2002433_11.jpg", + "2002433_12.jpg", + "2002433_13.jpg", + "2002433_14.jpg", + "2002433_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Simone Biles dancing with Serena Williams, on the giant piano, raining, hd, realistic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000370", + "images": [ + "4000370_00.jpg", + "4000370_01.jpg", + "4000370_02.jpg", + "4000370_03.jpg", + "4000370_04.jpg", + "4000370_05.jpg", + "4000370_06.jpg", + "4000370_07.jpg", + "4000370_08.jpg", + "4000370_09.jpg", + "4000370_10.jpg", + "4000370_11.jpg", + "4000370_12.jpg", + "4000370_13.jpg", + "4000370_14.jpg", + "4000370_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"one man dressed in a white sheet walks with a staff across the field, the rear view is far away, there is only one person with a staff in the frame \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "3002568", + "images": [ + "3002568_00.jpg", + "3002568_01.jpg", + "3002568_02.jpg", + "3002568_03.jpg", + "3002568_04.jpg", + "3002568_05.jpg", + "3002568_06.jpg", + "3002568_07.jpg", + "3002568_08.jpg", + "3002568_09.jpg", + "3002568_10.jpg", + "3002568_11.jpg", + "3002568_12.jpg", + "3002568_13.jpg", + "3002568_14.jpg", + "3002568_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"cyberpunk 3D Message: WildCard (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "0002588", + "images": [ + "0002588_00.jpg", + "0002588_01.jpg", + "0002588_02.jpg", + "0002588_03.jpg", + "0002588_04.jpg", + "0002588_05.jpg", + "0002588_06.jpg", + "0002588_07.jpg", + "0002588_08.jpg", + "0002588_09.jpg", + "0002588_10.jpg", + "0002588_11.jpg", + "0002588_12.jpg", + "0002588_13.jpg", + "0002588_14.jpg", + "0002588_15.jpg", + "0002588_16.jpg", + "0002588_17.jpg", + "0002588_18.jpg", + "0002588_19.jpg", + "0002588_20.jpg", + "0002588_21.jpg", + "0002588_22.jpg", + "0002588_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"have him walking through foggy village \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000507", + "images": [ + "4000507_00.jpg", + "4000507_01.jpg", + "4000507_02.jpg", + "4000507_03.jpg", + "4000507_04.jpg", + "4000507_05.jpg", + "4000507_06.jpg", + "4000507_07.jpg", + "4000507_08.jpg", + "4000507_09.jpg", + "4000507_10.jpg", + "4000507_11.jpg", + "4000507_12.jpg", + "4000507_13.jpg", + "4000507_14.jpg", + "4000507_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"H.R. Giger Biomechanics landscape; Happy and Surreal \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001130", + "images": [ + "6001130_00.jpg", + "6001130_01.jpg", + "6001130_02.jpg", + "6001130_03.jpg", + "6001130_04.jpg", + "6001130_05.jpg", + "6001130_06.jpg", + "6001130_07.jpg", + "6001130_08.jpg", + "6001130_09.jpg", + "6001130_10.jpg", + "6001130_11.jpg", + "6001130_12.jpg", + "6001130_13.jpg", + "6001130_14.jpg", + "6001130_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a scenery of a usa based company with a cool weather \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "5000852", + "images": [ + "5000852_00.jpg", + "5000852_01.jpg", + "5000852_02.jpg", + "5000852_03.jpg", + "5000852_04.jpg", + "5000852_05.jpg", + "5000852_06.jpg", + "5000852_07.jpg", + "5000852_08.jpg", + "5000852_09.jpg", + "5000852_10.jpg", + "5000852_11.jpg", + "5000852_12.jpg", + "5000852_13.jpg", + "5000852_14.jpg", + "5000852_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"the heroes have a unique and adventurous experience which in 3d \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "6000330", + "images": [ + "6000330_00.jpg", + "6000330_01.jpg", + "6000330_02.jpg", + "6000330_03.jpg", + "6000330_04.jpg", + "6000330_05.jpg", + "6000330_06.jpg", + "6000330_07.jpg", + "6000330_08.jpg", + "6000330_09.jpg", + "6000330_10.jpg", + "6000330_11.jpg", + "6000330_12.jpg", + "6000330_13.jpg", + "6000330_14.jpg", + "6000330_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a golden CHARIOT with seven white horses, flying CHARIOT in the sky, Photorealistic, Realistic, cinematic, 16:9, Ultra HD, \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 2\n" + } + ] + }, + { + "id": "3001890", + "images": [ + "3001890_00.jpg", + "3001890_01.jpg", + "3001890_02.jpg", + "3001890_03.jpg", + "3001890_04.jpg", + "3001890_05.jpg", + "3001890_06.jpg", + "3001890_07.jpg", + "3001890_08.jpg", + "3001890_09.jpg", + "3001890_10.jpg", + "3001890_11.jpg", + "3001890_12.jpg", + "3001890_13.jpg", + "3001890_14.jpg", + "3001890_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Be a little match girl, helpless video \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4001080", + "images": [ + "4001080_00.jpg", + "4001080_01.jpg", + "4001080_02.jpg", + "4001080_03.jpg", + "4001080_04.jpg", + "4001080_05.jpg", + "4001080_06.jpg", + "4001080_07.jpg", + "4001080_08.jpg", + "4001080_09.jpg", + "4001080_10.jpg", + "4001080_11.jpg", + "4001080_12.jpg", + "4001080_13.jpg", + "4001080_14.jpg", + "4001080_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"An illustration that portrays Daniel refusing to compromise his values, even in the face of outside pressures. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001416", + "images": [ + "6001416_00.jpg", + "6001416_01.jpg", + "6001416_02.jpg", + "6001416_03.jpg", + "6001416_04.jpg", + "6001416_05.jpg", + "6001416_06.jpg", + "6001416_07.jpg", + "6001416_08.jpg", + "6001416_09.jpg", + "6001416_10.jpg", + "6001416_11.jpg", + "6001416_12.jpg", + "6001416_13.jpg", + "6001416_14.jpg", + "6001416_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a cartoon animated woman working on the laptop experiencing an earthquake \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0002928", + "images": [ + "0002928_00.jpg", + "0002928_01.jpg", + "0002928_02.jpg", + "0002928_03.jpg", + "0002928_04.jpg", + "0002928_05.jpg", + "0002928_06.jpg", + "0002928_07.jpg", + "0002928_08.jpg", + "0002928_09.jpg", + "0002928_10.jpg", + "0002928_11.jpg", + "0002928_12.jpg", + "0002928_13.jpg", + "0002928_14.jpg", + "0002928_15.jpg", + "0002928_16.jpg", + "0002928_17.jpg", + "0002928_18.jpg", + "0002928_19.jpg", + "0002928_20.jpg", + "0002928_21.jpg", + "0002928_22.jpg", + "0002928_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"the detective decides to investigate the murder to find out who killed Julia and why. Julia continues to follow him, becoming his literal shadow. noir aesthetics \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002640", + "images": [ + "1002640_00.jpg", + "1002640_01.jpg", + "1002640_02.jpg", + "1002640_03.jpg", + "1002640_04.jpg", + "1002640_05.jpg", + "1002640_06.jpg", + "1002640_07.jpg", + "1002640_08.jpg", + "1002640_09.jpg", + "1002640_10.jpg", + "1002640_11.jpg", + "1002640_12.jpg", + "1002640_13.jpg", + "1002640_14.jpg", + "1002640_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"one girl toking and boy eating 2d,cartoon style, 4k \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001528", + "images": [ + "4001528_00.jpg", + "4001528_01.jpg", + "4001528_02.jpg", + "4001528_03.jpg", + "4001528_04.jpg", + "4001528_05.jpg", + "4001528_06.jpg", + "4001528_07.jpg", + "4001528_08.jpg", + "4001528_09.jpg", + "4001528_10.jpg", + "4001528_11.jpg", + "4001528_12.jpg", + "4001528_13.jpg", + "4001528_14.jpg", + "4001528_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"one Double meat burger falling onto a rustic wooden board in a dark yet warm ambiance illuminated by fire. format 9:16 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001986", + "images": [ + "7001986_00.jpg", + "7001986_01.jpg", + "7001986_02.jpg", + "7001986_03.jpg", + "7001986_04.jpg", + "7001986_05.jpg", + "7001986_06.jpg", + "7001986_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"plants tangled and intertwined with each other, super realistic, in the dry earth \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001578", + "images": [ + "1001578_00.jpg", + "1001578_01.jpg", + "1001578_02.jpg", + "1001578_03.jpg", + "1001578_04.jpg", + "1001578_05.jpg", + "1001578_06.jpg", + "1001578_07.jpg", + "1001578_08.jpg", + "1001578_09.jpg", + "1001578_10.jpg", + "1001578_11.jpg", + "1001578_12.jpg", + "1001578_13.jpg", + "1001578_14.jpg", + "1001578_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Sunset on the roof of the building, the girl stood watching, her hair flying. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000651", + "images": [ + "7000651_00.jpg", + "7000651_01.jpg", + "7000651_02.jpg", + "7000651_03.jpg", + "7000651_04.jpg", + "7000651_05.jpg", + "7000651_06.jpg", + "7000651_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Stormtroopers welcome on a 4th of July celebration \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2001977", + "images": [ + "2001977_00.jpg", + "2001977_01.jpg", + "2001977_02.jpg", + "2001977_03.jpg", + "2001977_04.jpg", + "2001977_05.jpg", + "2001977_06.jpg", + "2001977_07.jpg", + "2001977_08.jpg", + "2001977_09.jpg", + "2001977_10.jpg", + "2001977_11.jpg", + "2001977_12.jpg", + "2001977_13.jpg", + "2001977_14.jpg", + "2001977_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Snow White in her castle, where she is loved by all but envied by her stepmother, the Queen. 9:16 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001798", + "images": [ + "1001798_00.jpg", + "1001798_01.jpg", + "1001798_02.jpg", + "1001798_03.jpg", + "1001798_04.jpg", + "1001798_05.jpg", + "1001798_06.jpg", + "1001798_07.jpg", + "1001798_08.jpg", + "1001798_09.jpg", + "1001798_10.jpg", + "1001798_11.jpg", + "1001798_12.jpg", + "1001798_13.jpg", + "1001798_14.jpg", + "1001798_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A group of gangs are gathering in a luxury house to drink. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "4000553", + "images": [ + "4000553_00.jpg", + "4000553_01.jpg", + "4000553_02.jpg", + "4000553_03.jpg", + "4000553_04.jpg", + "4000553_05.jpg", + "4000553_06.jpg", + "4000553_07.jpg", + "4000553_08.jpg", + "4000553_09.jpg", + "4000553_10.jpg", + "4000553_11.jpg", + "4000553_12.jpg", + "4000553_13.jpg", + "4000553_14.jpg", + "4000553_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Wawa Gas Station Message: NK Packs Fudge (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001272", + "images": [ + "1001272_00.jpg", + "1001272_01.jpg", + "1001272_02.jpg", + "1001272_03.jpg", + "1001272_04.jpg", + "1001272_05.jpg", + "1001272_06.jpg", + "1001272_07.jpg", + "1001272_08.jpg", + "1001272_09.jpg", + "1001272_10.jpg", + "1001272_11.jpg", + "1001272_12.jpg", + "1001272_13.jpg", + "1001272_14.jpg", + "1001272_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"on an old 1950 tv set, a woman dressed as catrina walking through the cemetery at night with a fog, super realistic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002634", + "images": [ + "3002634_00.jpg", + "3002634_01.jpg", + "3002634_02.jpg", + "3002634_03.jpg", + "3002634_04.jpg", + "3002634_05.jpg", + "3002634_06.jpg", + "3002634_07.jpg", + "3002634_08.jpg", + "3002634_09.jpg", + "3002634_10.jpg", + "3002634_11.jpg", + "3002634_12.jpg", + "3002634_13.jpg", + "3002634_14.jpg", + "3002634_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"During the Jurassic period, a mammoth was running in a realistic style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "7000705", + "images": [ + "7000705_00.jpg", + "7000705_01.jpg", + "7000705_02.jpg", + "7000705_03.jpg", + "7000705_04.jpg", + "7000705_05.jpg", + "7000705_06.jpg", + "7000705_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A white bird flying on a green screen \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "4001530", + "images": [ + "4001530_00.jpg", + "4001530_01.jpg", + "4001530_02.jpg", + "4001530_03.jpg", + "4001530_04.jpg", + "4001530_05.jpg", + "4001530_06.jpg", + "4001530_07.jpg", + "4001530_08.jpg", + "4001530_09.jpg", + "4001530_10.jpg", + "4001530_11.jpg", + "4001530_12.jpg", + "4001530_13.jpg", + "4001530_14.jpg", + "4001530_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Sequence of the tragic events of November 18, including the assassination of Congressman Ryan. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000471", + "images": [ + "6000471_00.jpg", + "6000471_01.jpg", + "6000471_02.jpg", + "6000471_03.jpg", + "6000471_04.jpg", + "6000471_05.jpg", + "6000471_06.jpg", + "6000471_07.jpg", + "6000471_08.jpg", + "6000471_09.jpg", + "6000471_10.jpg", + "6000471_11.jpg", + "6000471_12.jpg", + "6000471_13.jpg", + "6000471_14.jpg", + "6000471_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"marching army at time square like futuristic post apocalyptic city, highly detailed, 4K, realistic, hyper realistic, 35 mm lens, good anatomy, neon lights, good lights, excellent photgraphy, cinematic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "2001202", + "images": [ + "2001202_00.jpg", + "2001202_01.jpg", + "2001202_02.jpg", + "2001202_03.jpg", + "2001202_04.jpg", + "2001202_05.jpg", + "2001202_06.jpg", + "2001202_07.jpg", + "2001202_08.jpg", + "2001202_09.jpg", + "2001202_10.jpg", + "2001202_11.jpg", + "2001202_12.jpg", + "2001202_13.jpg", + "2001202_14.jpg", + "2001202_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Describe the ancient haveli in Kolkata where the priceless clock is kept. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000688", + "images": [ + "3000688_00.jpg", + "3000688_01.jpg", + "3000688_02.jpg", + "3000688_03.jpg", + "3000688_04.jpg", + "3000688_05.jpg", + "3000688_06.jpg", + "3000688_07.jpg", + "3000688_08.jpg", + "3000688_09.jpg", + "3000688_10.jpg", + "3000688_11.jpg", + "3000688_12.jpg", + "3000688_13.jpg", + "3000688_14.jpg", + "3000688_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Generate an image of the wise advisers discussing their predictions with a indian King Suddhodana, highlighting the seriousness of the foreseen events that might lead prince to seek a different path. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "7001711", + "images": [ + "7001711_00.jpg", + "7001711_01.jpg", + "7001711_02.jpg", + "7001711_03.jpg", + "7001711_04.jpg", + "7001711_05.jpg", + "7001711_06.jpg", + "7001711_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"seikh climbing the hill fastly to get the got back \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000388", + "images": [ + "6000388_00.jpg", + "6000388_01.jpg", + "6000388_02.jpg", + "6000388_03.jpg", + "6000388_04.jpg", + "6000388_05.jpg", + "6000388_06.jpg", + "6000388_07.jpg", + "6000388_08.jpg", + "6000388_09.jpg", + "6000388_10.jpg", + "6000388_11.jpg", + "6000388_12.jpg", + "6000388_13.jpg", + "6000388_14.jpg", + "6000388_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a scattering of sparkling precious stones of different colors silver gold red green blue yellow Message: 15 years (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "1002603", + "images": [ + "1002603_00.jpg", + "1002603_01.jpg", + "1002603_02.jpg", + "1002603_03.jpg", + "1002603_04.jpg", + "1002603_05.jpg", + "1002603_06.jpg", + "1002603_07.jpg", + "1002603_08.jpg", + "1002603_09.jpg", + "1002603_10.jpg", + "1002603_11.jpg", + "1002603_12.jpg", + "1002603_13.jpg", + "1002603_14.jpg", + "1002603_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"make me a 400x400 photo of vietnam national day content. Holiday period from September 1, 2023 to September 3, 2023 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002829", + "images": [ + "2002829_00.jpg", + "2002829_01.jpg", + "2002829_02.jpg", + "2002829_03.jpg", + "2002829_04.jpg", + "2002829_05.jpg", + "2002829_06.jpg", + "2002829_07.jpg", + "2002829_08.jpg", + "2002829_09.jpg", + "2002829_10.jpg", + "2002829_11.jpg", + "2002829_12.jpg", + "2002829_13.jpg", + "2002829_14.jpg", + "2002829_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"virat kohli dancing on the beach with baber azam \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002468", + "images": [ + "2002468_00.jpg", + "2002468_01.jpg", + "2002468_02.jpg", + "2002468_03.jpg", + "2002468_04.jpg", + "2002468_05.jpg", + "2002468_06.jpg", + "2002468_07.jpg", + "2002468_08.jpg", + "2002468_09.jpg", + "2002468_10.jpg", + "2002468_11.jpg", + "2002468_12.jpg", + "2002468_13.jpg", + "2002468_14.jpg", + "2002468_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"and the walls were lined with artwork from the previous school years. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1000835", + "images": [ + "1000835_00.jpg", + "1000835_01.jpg", + "1000835_02.jpg", + "1000835_03.jpg", + "1000835_04.jpg", + "1000835_05.jpg", + "1000835_06.jpg", + "1000835_07.jpg", + "1000835_08.jpg", + "1000835_09.jpg", + "1000835_10.jpg", + "1000835_11.jpg", + "1000835_12.jpg", + "1000835_13.jpg", + "1000835_14.jpg", + "1000835_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"3d cartoon in a cave, one pirate pushed other down in the lava \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1000836", + "images": [ + "1000836_00.jpg", + "1000836_01.jpg", + "1000836_02.jpg", + "1000836_03.jpg", + "1000836_04.jpg", + "1000836_05.jpg", + "1000836_06.jpg", + "1000836_07.jpg", + "1000836_08.jpg", + "1000836_09.jpg", + "1000836_10.jpg", + "1000836_11.jpg", + "1000836_12.jpg", + "1000836_13.jpg", + "1000836_14.jpg", + "1000836_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a old merchant giving money to his teenagers boys \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2001261", + "images": [ + "2001261_00.jpg", + "2001261_01.jpg", + "2001261_02.jpg", + "2001261_03.jpg", + "2001261_04.jpg", + "2001261_05.jpg", + "2001261_06.jpg", + "2001261_07.jpg", + "2001261_08.jpg", + "2001261_09.jpg", + "2001261_10.jpg", + "2001261_11.jpg", + "2001261_12.jpg", + "2001261_13.jpg", + "2001261_14.jpg", + "2001261_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A tiger is taking a bath \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002748", + "images": [ + "1002748_00.jpg", + "1002748_01.jpg", + "1002748_02.jpg", + "1002748_03.jpg", + "1002748_04.jpg", + "1002748_05.jpg", + "1002748_06.jpg", + "1002748_07.jpg", + "1002748_08.jpg", + "1002748_09.jpg", + "1002748_10.jpg", + "1002748_11.jpg", + "1002748_12.jpg", + "1002748_13.jpg", + "1002748_14.jpg", + "1002748_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Harry Potter shakes hands with Voldemort in the Miami Ghetto in GTA \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002265", + "images": [ + "3002265_00.jpg", + "3002265_01.jpg", + "3002265_02.jpg", + "3002265_03.jpg", + "3002265_04.jpg", + "3002265_05.jpg", + "3002265_06.jpg", + "3002265_07.jpg", + "3002265_08.jpg", + "3002265_09.jpg", + "3002265_10.jpg", + "3002265_11.jpg", + "3002265_12.jpg", + "3002265_13.jpg", + "3002265_14.jpg", + "3002265_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"hyperrealistic, light painting, writing Message: AMAZING (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "0002612", + "images": [ + "0002612_00.jpg", + "0002612_01.jpg", + "0002612_02.jpg", + "0002612_03.jpg", + "0002612_04.jpg", + "0002612_05.jpg", + "0002612_06.jpg", + "0002612_07.jpg", + "0002612_08.jpg", + "0002612_09.jpg", + "0002612_10.jpg", + "0002612_11.jpg", + "0002612_12.jpg", + "0002612_13.jpg", + "0002612_14.jpg", + "0002612_15.jpg", + "0002612_16.jpg", + "0002612_17.jpg", + "0002612_18.jpg", + "0002612_19.jpg", + "0002612_20.jpg", + "0002612_21.jpg", + "0002612_22.jpg", + "0002612_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"anime stars spreading across the universe timelapse \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0000781", + "images": [ + "0000781_00.jpg", + "0000781_01.jpg", + "0000781_02.jpg", + "0000781_03.jpg", + "0000781_04.jpg", + "0000781_05.jpg", + "0000781_06.jpg", + "0000781_07.jpg", + "0000781_08.jpg", + "0000781_09.jpg", + "0000781_10.jpg", + "0000781_11.jpg", + "0000781_12.jpg", + "0000781_13.jpg", + "0000781_14.jpg", + "0000781_15.jpg", + "0000781_16.jpg", + "0000781_17.jpg", + "0000781_18.jpg", + "0000781_19.jpg", + "0000781_20.jpg", + "0000781_21.jpg", + "0000781_22.jpg", + "0000781_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"We see the haunting, surreal image of Cthulhu, his eerie, otherworldly presence floating in the void of space, passing by sinister and ominous planets. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001543", + "images": [ + "7001543_00.jpg", + "7001543_01.jpg", + "7001543_02.jpg", + "7001543_03.jpg", + "7001543_04.jpg", + "7001543_05.jpg", + "7001543_06.jpg", + "7001543_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a robot teaching how to reppair something to a kid \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "8000761", + "images": [ + "8000761_00.jpg", + "8000761_01.jpg", + "8000761_02.jpg", + "8000761_03.jpg", + "8000761_04.jpg", + "8000761_05.jpg", + "8000761_06.jpg", + "8000761_07.jpg", + "8000761_08.jpg", + "8000761_09.jpg", + "8000761_10.jpg", + "8000761_11.jpg", + "8000761_12.jpg", + "8000761_13.jpg", + "8000761_14.jpg", + "8000761_15.jpg", + "8000761_16.jpg", + "8000761_17.jpg", + "8000761_18.jpg", + "8000761_19.jpg", + "8000761_20.jpg", + "8000761_21.jpg", + "8000761_22.jpg", + "8000761_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"women wearing glasses audrey hepburn style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2000522", + "images": [ + "2000522_00.jpg", + "2000522_01.jpg", + "2000522_02.jpg", + "2000522_03.jpg", + "2000522_04.jpg", + "2000522_05.jpg", + "2000522_06.jpg", + "2000522_07.jpg", + "2000522_08.jpg", + "2000522_09.jpg", + "2000522_10.jpg", + "2000522_11.jpg", + "2000522_12.jpg", + "2000522_13.jpg", + "2000522_14.jpg", + "2000522_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Town in winter with house panorama \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000095", + "images": [ + "3000095_00.jpg", + "3000095_01.jpg", + "3000095_02.jpg", + "3000095_03.jpg", + "3000095_04.jpg", + "3000095_05.jpg", + "3000095_06.jpg", + "3000095_07.jpg", + "3000095_08.jpg", + "3000095_09.jpg", + "3000095_10.jpg", + "3000095_11.jpg", + "3000095_12.jpg", + "3000095_13.jpg", + "3000095_14.jpg", + "3000095_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Only one cat is on the roof. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001310", + "images": [ + "6001310_00.jpg", + "6001310_01.jpg", + "6001310_02.jpg", + "6001310_03.jpg", + "6001310_04.jpg", + "6001310_05.jpg", + "6001310_06.jpg", + "6001310_07.jpg", + "6001310_08.jpg", + "6001310_09.jpg", + "6001310_10.jpg", + "6001310_11.jpg", + "6001310_12.jpg", + "6001310_13.jpg", + "6001310_14.jpg", + "6001310_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"I need a promotional video for a clothing store, where several animated items of clothing appear \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001637", + "images": [ + "4001637_00.jpg", + "4001637_01.jpg", + "4001637_02.jpg", + "4001637_03.jpg", + "4001637_04.jpg", + "4001637_05.jpg", + "4001637_06.jpg", + "4001637_07.jpg", + "4001637_08.jpg", + "4001637_09.jpg", + "4001637_10.jpg", + "4001637_11.jpg", + "4001637_12.jpg", + "4001637_13.jpg", + "4001637_14.jpg", + "4001637_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"minecraft mountain range with two characters looking in the distance \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001699", + "images": [ + "6001699_00.jpg", + "6001699_01.jpg", + "6001699_02.jpg", + "6001699_03.jpg", + "6001699_04.jpg", + "6001699_05.jpg", + "6001699_06.jpg", + "6001699_07.jpg", + "6001699_08.jpg", + "6001699_09.jpg", + "6001699_10.jpg", + "6001699_11.jpg", + "6001699_12.jpg", + "6001699_13.jpg", + "6001699_14.jpg", + "6001699_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"sharp focus, high detailed, Sakyamuni is sitting in meditation, 4K, HDR \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000658", + "images": [ + "6000658_00.jpg", + "6000658_01.jpg", + "6000658_02.jpg", + "6000658_03.jpg", + "6000658_04.jpg", + "6000658_05.jpg", + "6000658_06.jpg", + "6000658_07.jpg", + "6000658_08.jpg", + "6000658_09.jpg", + "6000658_10.jpg", + "6000658_11.jpg", + "6000658_12.jpg", + "6000658_13.jpg", + "6000658_14.jpg", + "6000658_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A cross with a piece of red cloth wrapped around it waving in the wind during a sunrise in a beautiful hilly environment with waterfalls and birds flying in the sky \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "0000145", + "images": [ + "0000145_00.jpg", + "0000145_01.jpg", + "0000145_02.jpg", + "0000145_03.jpg", + "0000145_04.jpg", + "0000145_05.jpg", + "0000145_06.jpg", + "0000145_07.jpg", + "0000145_08.jpg", + "0000145_09.jpg", + "0000145_10.jpg", + "0000145_11.jpg", + "0000145_12.jpg", + "0000145_13.jpg", + "0000145_14.jpg", + "0000145_15.jpg", + "0000145_16.jpg", + "0000145_17.jpg", + "0000145_18.jpg", + "0000145_19.jpg", + "0000145_20.jpg", + "0000145_21.jpg", + "0000145_22.jpg", + "0000145_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"wide angle from the balcony: looking the landscape with Giger palaces in ruins, satanic monuments, totems of evil creatures, dramatic cloudy sky, mystical, hyperdetailed, intricate perfect details, complex, horror art. Illustration by Dan Seagrave \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001669", + "images": [ + "1001669_00.jpg", + "1001669_01.jpg", + "1001669_02.jpg", + "1001669_03.jpg", + "1001669_04.jpg", + "1001669_05.jpg", + "1001669_06.jpg", + "1001669_07.jpg", + "1001669_08.jpg", + "1001669_09.jpg", + "1001669_10.jpg", + "1001669_11.jpg", + "1001669_12.jpg", + "1001669_13.jpg", + "1001669_14.jpg", + "1001669_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"beautiful birds flying in the forest at twilight, lens flare \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000397", + "images": [ + "3000397_00.jpg", + "3000397_01.jpg", + "3000397_02.jpg", + "3000397_03.jpg", + "3000397_04.jpg", + "3000397_05.jpg", + "3000397_06.jpg", + "3000397_07.jpg", + "3000397_08.jpg", + "3000397_09.jpg", + "3000397_10.jpg", + "3000397_11.jpg", + "3000397_12.jpg", + "3000397_13.jpg", + "3000397_14.jpg", + "3000397_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a renaissance dressed women is walkin down a palace room \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001056", + "images": [ + "4001056_00.jpg", + "4001056_01.jpg", + "4001056_02.jpg", + "4001056_03.jpg", + "4001056_04.jpg", + "4001056_05.jpg", + "4001056_06.jpg", + "4001056_07.jpg", + "4001056_08.jpg", + "4001056_09.jpg", + "4001056_10.jpg", + "4001056_11.jpg", + "4001056_12.jpg", + "4001056_13.jpg", + "4001056_14.jpg", + "4001056_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"an electric robot in the shape of a fish sucks water under pressure, transports water along a conveyor to a special department and splashes water into the river through a carbon filter \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "0001697", + "images": [ + "0001697_00.jpg", + "0001697_01.jpg", + "0001697_02.jpg", + "0001697_03.jpg", + "0001697_04.jpg", + "0001697_05.jpg", + "0001697_06.jpg", + "0001697_07.jpg", + "0001697_08.jpg", + "0001697_09.jpg", + "0001697_10.jpg", + "0001697_11.jpg", + "0001697_12.jpg", + "0001697_13.jpg", + "0001697_14.jpg", + "0001697_15.jpg", + "0001697_16.jpg", + "0001697_17.jpg", + "0001697_18.jpg", + "0001697_19.jpg", + "0001697_20.jpg", + "0001697_21.jpg", + "0001697_22.jpg", + "0001697_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Dawn over the sea. Birds fly by. Light splashing of water on small coastal pebbles. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000990", + "images": [ + "5000990_00.jpg", + "5000990_01.jpg", + "5000990_02.jpg", + "5000990_03.jpg", + "5000990_04.jpg", + "5000990_05.jpg", + "5000990_06.jpg", + "5000990_07.jpg", + "5000990_08.jpg", + "5000990_09.jpg", + "5000990_10.jpg", + "5000990_11.jpg", + "5000990_12.jpg", + "5000990_13.jpg", + "5000990_14.jpg", + "5000990_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"people praying a statue of Ronald McDonald in the church of McDonalds \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "6001774", + "images": [ + "6001774_00.jpg", + "6001774_01.jpg", + "6001774_02.jpg", + "6001774_03.jpg", + "6001774_04.jpg", + "6001774_05.jpg", + "6001774_06.jpg", + "6001774_07.jpg", + "6001774_08.jpg", + "6001774_09.jpg", + "6001774_10.jpg", + "6001774_11.jpg", + "6001774_12.jpg", + "6001774_13.jpg", + "6001774_14.jpg", + "6001774_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"remove all out of focus pixels \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2000516", + "images": [ + "2000516_00.jpg", + "2000516_01.jpg", + "2000516_02.jpg", + "2000516_03.jpg", + "2000516_04.jpg", + "2000516_05.jpg", + "2000516_06.jpg", + "2000516_07.jpg", + "2000516_08.jpg", + "2000516_09.jpg", + "2000516_10.jpg", + "2000516_11.jpg", + "2000516_12.jpg", + "2000516_13.jpg", + "2000516_14.jpg", + "2000516_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"cartoon bear in river watching salmon \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "0000026", + "images": [ + "0000026_00.jpg", + "0000026_01.jpg", + "0000026_02.jpg", + "0000026_03.jpg", + "0000026_04.jpg", + "0000026_05.jpg", + "0000026_06.jpg", + "0000026_07.jpg", + "0000026_08.jpg", + "0000026_09.jpg", + "0000026_10.jpg", + "0000026_11.jpg", + "0000026_12.jpg", + "0000026_13.jpg", + "0000026_14.jpg", + "0000026_15.jpg", + "0000026_16.jpg", + "0000026_17.jpg", + "0000026_18.jpg", + "0000026_19.jpg", + "0000026_20.jpg", + "0000026_21.jpg", + "0000026_22.jpg", + "0000026_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"tornado throwing an elephant and animals in the air, violent storm, ar:9:16 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "2000623", + "images": [ + "2000623_00.jpg", + "2000623_01.jpg", + "2000623_02.jpg", + "2000623_03.jpg", + "2000623_04.jpg", + "2000623_05.jpg", + "2000623_06.jpg", + "2000623_07.jpg", + "2000623_08.jpg", + "2000623_09.jpg", + "2000623_10.jpg", + "2000623_11.jpg", + "2000623_12.jpg", + "2000623_13.jpg", + "2000623_14.jpg", + "2000623_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Sniper in 1944, adjusting his scope and preparing to take a shot, 24K, AR 16:9, cinematography, camera zoom in \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "2001102", + "images": [ + "2001102_00.jpg", + "2001102_01.jpg", + "2001102_02.jpg", + "2001102_03.jpg", + "2001102_04.jpg", + "2001102_05.jpg", + "2001102_06.jpg", + "2001102_07.jpg", + "2001102_08.jpg", + "2001102_09.jpg", + "2001102_10.jpg", + "2001102_11.jpg", + "2001102_12.jpg", + "2001102_13.jpg", + "2001102_14.jpg", + "2001102_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"flowers moving, breezy, frangipani flowers, blinking lights \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4001055", + "images": [ + "4001055_00.jpg", + "4001055_01.jpg", + "4001055_02.jpg", + "4001055_03.jpg", + "4001055_04.jpg", + "4001055_05.jpg", + "4001055_06.jpg", + "4001055_07.jpg", + "4001055_08.jpg", + "4001055_09.jpg", + "4001055_10.jpg", + "4001055_11.jpg", + "4001055_12.jpg", + "4001055_13.jpg", + "4001055_14.jpg", + "4001055_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"galaxy creation, how our universe is created \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2000980", + "images": [ + "2000980_00.jpg", + "2000980_01.jpg", + "2000980_02.jpg", + "2000980_03.jpg", + "2000980_04.jpg", + "2000980_05.jpg", + "2000980_06.jpg", + "2000980_07.jpg", + "2000980_08.jpg", + "2000980_09.jpg", + "2000980_10.jpg", + "2000980_11.jpg", + "2000980_12.jpg", + "2000980_13.jpg", + "2000980_14.jpg", + "2000980_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"slow and subltle movement and zoom in image \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "8000452", + "images": [ + "8000452_00.jpg", + "8000452_01.jpg", + "8000452_02.jpg", + "8000452_03.jpg", + "8000452_04.jpg", + "8000452_05.jpg", + "8000452_06.jpg", + "8000452_07.jpg", + "8000452_08.jpg", + "8000452_09.jpg", + "8000452_10.jpg", + "8000452_11.jpg", + "8000452_12.jpg", + "8000452_13.jpg", + "8000452_14.jpg", + "8000452_15.jpg", + "8000452_16.jpg", + "8000452_17.jpg", + "8000452_18.jpg", + "8000452_19.jpg", + "8000452_20.jpg", + "8000452_21.jpg", + "8000452_22.jpg", + "8000452_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"time dilation in high school girls \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002225", + "images": [ + "3002225_00.jpg", + "3002225_01.jpg", + "3002225_02.jpg", + "3002225_03.jpg", + "3002225_04.jpg", + "3002225_05.jpg", + "3002225_06.jpg", + "3002225_07.jpg", + "3002225_08.jpg", + "3002225_09.jpg", + "3002225_10.jpg", + "3002225_11.jpg", + "3002225_12.jpg", + "3002225_13.jpg", + "3002225_14.jpg", + "3002225_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A stunning view of the underwater city of Rapture, glimmering lights, and towering art deco architecture. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000540", + "images": [ + "6000540_00.jpg", + "6000540_01.jpg", + "6000540_02.jpg", + "6000540_03.jpg", + "6000540_04.jpg", + "6000540_05.jpg", + "6000540_06.jpg", + "6000540_07.jpg", + "6000540_08.jpg", + "6000540_09.jpg", + "6000540_10.jpg", + "6000540_11.jpg", + "6000540_12.jpg", + "6000540_13.jpg", + "6000540_14.jpg", + "6000540_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Into the Unknown: At the stroke of midnight, the detective arrives at a grand mansion on the outskirts of the city, its imposing silhouette looming against the starlit sky. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "5000624", + "images": [ + "5000624_00.jpg", + "5000624_01.jpg", + "5000624_02.jpg", + "5000624_03.jpg", + "5000624_04.jpg", + "5000624_05.jpg", + "5000624_06.jpg", + "5000624_07.jpg", + "5000624_08.jpg", + "5000624_09.jpg", + "5000624_10.jpg", + "5000624_11.jpg", + "5000624_12.jpg", + "5000624_13.jpg", + "5000624_14.jpg", + "5000624_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A Mongolian khan, wearing a luxurious robe, rides a warhorse, and commands his soldiers to advance. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "3000592", + "images": [ + "3000592_00.jpg", + "3000592_01.jpg", + "3000592_02.jpg", + "3000592_03.jpg", + "3000592_04.jpg", + "3000592_05.jpg", + "3000592_06.jpg", + "3000592_07.jpg", + "3000592_08.jpg", + "3000592_09.jpg", + "3000592_10.jpg", + "3000592_11.jpg", + "3000592_12.jpg", + "3000592_13.jpg", + "3000592_14.jpg", + "3000592_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"birds are flying in the sky in groups in a village. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "7001587", + "images": [ + "7001587_00.jpg", + "7001587_01.jpg", + "7001587_02.jpg", + "7001587_03.jpg", + "7001587_04.jpg", + "7001587_05.jpg", + "7001587_06.jpg", + "7001587_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A rotating protein structure for science research \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002892", + "images": [ + "2002892_00.jpg", + "2002892_01.jpg", + "2002892_02.jpg", + "2002892_03.jpg", + "2002892_04.jpg", + "2002892_05.jpg", + "2002892_06.jpg", + "2002892_07.jpg", + "2002892_08.jpg", + "2002892_09.jpg", + "2002892_10.jpg", + "2002892_11.jpg", + "2002892_12.jpg", + "2002892_13.jpg", + "2002892_14.jpg", + "2002892_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"several female human looking robots power down and fall \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000138", + "images": [ + "6000138_00.jpg", + "6000138_01.jpg", + "6000138_02.jpg", + "6000138_03.jpg", + "6000138_04.jpg", + "6000138_05.jpg", + "6000138_06.jpg", + "6000138_07.jpg", + "6000138_08.jpg", + "6000138_09.jpg", + "6000138_10.jpg", + "6000138_11.jpg", + "6000138_12.jpg", + "6000138_13.jpg", + "6000138_14.jpg", + "6000138_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"from snow capped mountains of himalaya to the triple seas meeting point at kanyakumuri you present a kaledioscope picture of sky punching mountains lakes and majestic rivers swift flowing streams and thunderstorm water flows \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "6000446", + "images": [ + "6000446_00.jpg", + "6000446_01.jpg", + "6000446_02.jpg", + "6000446_03.jpg", + "6000446_04.jpg", + "6000446_05.jpg", + "6000446_06.jpg", + "6000446_07.jpg", + "6000446_08.jpg", + "6000446_09.jpg", + "6000446_10.jpg", + "6000446_11.jpg", + "6000446_12.jpg", + "6000446_13.jpg", + "6000446_14.jpg", + "6000446_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"They watched in awe as the forgotten symphony wove through the air \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002336", + "images": [ + "2002336_00.jpg", + "2002336_01.jpg", + "2002336_02.jpg", + "2002336_03.jpg", + "2002336_04.jpg", + "2002336_05.jpg", + "2002336_06.jpg", + "2002336_07.jpg", + "2002336_08.jpg", + "2002336_09.jpg", + "2002336_10.jpg", + "2002336_11.jpg", + "2002336_12.jpg", + "2002336_13.jpg", + "2002336_14.jpg", + "2002336_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"ice crystals forming on a black bacgkround \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000307", + "images": [ + "6000307_00.jpg", + "6000307_01.jpg", + "6000307_02.jpg", + "6000307_03.jpg", + "6000307_04.jpg", + "6000307_05.jpg", + "6000307_06.jpg", + "6000307_07.jpg", + "6000307_08.jpg", + "6000307_09.jpg", + "6000307_10.jpg", + "6000307_11.jpg", + "6000307_12.jpg", + "6000307_13.jpg", + "6000307_14.jpg", + "6000307_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"make the puppy talk for 5 seconds \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002662", + "images": [ + "3002662_00.jpg", + "3002662_01.jpg", + "3002662_02.jpg", + "3002662_03.jpg", + "3002662_04.jpg", + "3002662_05.jpg", + "3002662_06.jpg", + "3002662_07.jpg", + "3002662_08.jpg", + "3002662_09.jpg", + "3002662_10.jpg", + "3002662_11.jpg", + "3002662_12.jpg", + "3002662_13.jpg", + "3002662_14.jpg", + "3002662_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a pot of ramen cooling down \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000287", + "images": [ + "3000287_00.jpg", + "3000287_01.jpg", + "3000287_02.jpg", + "3000287_03.jpg", + "3000287_04.jpg", + "3000287_05.jpg", + "3000287_06.jpg", + "3000287_07.jpg", + "3000287_08.jpg", + "3000287_09.jpg", + "3000287_10.jpg", + "3000287_11.jpg", + "3000287_12.jpg", + "3000287_13.jpg", + "3000287_14.jpg", + "3000287_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"fighting two super heroes from ancient civilization, 8k realistic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "6001661", + "images": [ + "6001661_00.jpg", + "6001661_01.jpg", + "6001661_02.jpg", + "6001661_03.jpg", + "6001661_04.jpg", + "6001661_05.jpg", + "6001661_06.jpg", + "6001661_07.jpg", + "6001661_08.jpg", + "6001661_09.jpg", + "6001661_10.jpg", + "6001661_11.jpg", + "6001661_12.jpg", + "6001661_13.jpg", + "6001661_14.jpg", + "6001661_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"statue of liberty flying huge space ship over mountains \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "7001977", + "images": [ + "7001977_00.jpg", + "7001977_01.jpg", + "7001977_02.jpg", + "7001977_03.jpg", + "7001977_04.jpg", + "7001977_05.jpg", + "7001977_06.jpg", + "7001977_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"acute beautiful girl in forest Disney style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002376", + "images": [ + "2002376_00.jpg", + "2002376_01.jpg", + "2002376_02.jpg", + "2002376_03.jpg", + "2002376_04.jpg", + "2002376_05.jpg", + "2002376_06.jpg", + "2002376_07.jpg", + "2002376_08.jpg", + "2002376_09.jpg", + "2002376_10.jpg", + "2002376_11.jpg", + "2002376_12.jpg", + "2002376_13.jpg", + "2002376_14.jpg", + "2002376_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Worms eye view of ants running around food on a kitchen table, macro \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000229", + "images": [ + "3000229_00.jpg", + "3000229_01.jpg", + "3000229_02.jpg", + "3000229_03.jpg", + "3000229_04.jpg", + "3000229_05.jpg", + "3000229_06.jpg", + "3000229_07.jpg", + "3000229_08.jpg", + "3000229_09.jpg", + "3000229_10.jpg", + "3000229_11.jpg", + "3000229_12.jpg", + "3000229_13.jpg", + "3000229_14.jpg", + "3000229_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a roller skate in the street \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "0002235", + "images": [ + "0002235_00.jpg", + "0002235_01.jpg", + "0002235_02.jpg", + "0002235_03.jpg", + "0002235_04.jpg", + "0002235_05.jpg", + "0002235_06.jpg", + "0002235_07.jpg", + "0002235_08.jpg", + "0002235_09.jpg", + "0002235_10.jpg", + "0002235_11.jpg", + "0002235_12.jpg", + "0002235_13.jpg", + "0002235_14.jpg", + "0002235_15.jpg", + "0002235_16.jpg", + "0002235_17.jpg", + "0002235_18.jpg", + "0002235_19.jpg", + "0002235_20.jpg", + "0002235_21.jpg", + "0002235_22.jpg", + "0002235_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Dwarfs in dwarf costumes on a forest road smoke and meet a bus \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3000649", + "images": [ + "3000649_00.jpg", + "3000649_01.jpg", + "3000649_02.jpg", + "3000649_03.jpg", + "3000649_04.jpg", + "3000649_05.jpg", + "3000649_06.jpg", + "3000649_07.jpg", + "3000649_08.jpg", + "3000649_09.jpg", + "3000649_10.jpg", + "3000649_11.jpg", + "3000649_12.jpg", + "3000649_13.jpg", + "3000649_14.jpg", + "3000649_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a space rocket with an alien inside Message: imposter (Font: MODERN) \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "2000176", + "images": [ + "2000176_00.jpg", + "2000176_01.jpg", + "2000176_02.jpg", + "2000176_03.jpg", + "2000176_04.jpg", + "2000176_05.jpg", + "2000176_06.jpg", + "2000176_07.jpg", + "2000176_08.jpg", + "2000176_09.jpg", + "2000176_10.jpg", + "2000176_11.jpg", + "2000176_12.jpg", + "2000176_13.jpg", + "2000176_14.jpg", + "2000176_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a video like a view from the camera going into the deep ocean, you can see bubbles of air on the side \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001459", + "images": [ + "6001459_00.jpg", + "6001459_01.jpg", + "6001459_02.jpg", + "6001459_03.jpg", + "6001459_04.jpg", + "6001459_05.jpg", + "6001459_06.jpg", + "6001459_07.jpg", + "6001459_08.jpg", + "6001459_09.jpg", + "6001459_10.jpg", + "6001459_11.jpg", + "6001459_12.jpg", + "6001459_13.jpg", + "6001459_14.jpg", + "6001459_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"man posing in front of lake with fish jumping out of the water \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1000109", + "images": [ + "1000109_00.jpg", + "1000109_01.jpg", + "1000109_02.jpg", + "1000109_03.jpg", + "1000109_04.jpg", + "1000109_05.jpg", + "1000109_06.jpg", + "1000109_07.jpg", + "1000109_08.jpg", + "1000109_09.jpg", + "1000109_10.jpg", + "1000109_11.jpg", + "1000109_12.jpg", + "1000109_13.jpg", + "1000109_14.jpg", + "1000109_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a fish in sky with jet \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002811", + "images": [ + "3002811_00.jpg", + "3002811_01.jpg", + "3002811_02.jpg", + "3002811_03.jpg", + "3002811_04.jpg", + "3002811_05.jpg", + "3002811_06.jpg", + "3002811_07.jpg", + "3002811_08.jpg", + "3002811_09.jpg", + "3002811_10.jpg", + "3002811_11.jpg", + "3002811_12.jpg", + "3002811_13.jpg", + "3002811_14.jpg", + "3002811_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"pigeon in a skirt dancing ballet in anime style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "0000652", + "images": [ + "0000652_00.jpg", + "0000652_01.jpg", + "0000652_02.jpg", + "0000652_03.jpg", + "0000652_04.jpg", + "0000652_05.jpg", + "0000652_06.jpg", + "0000652_07.jpg", + "0000652_08.jpg", + "0000652_09.jpg", + "0000652_10.jpg", + "0000652_11.jpg", + "0000652_12.jpg", + "0000652_13.jpg", + "0000652_14.jpg", + "0000652_15.jpg", + "0000652_16.jpg", + "0000652_17.jpg", + "0000652_18.jpg", + "0000652_19.jpg", + "0000652_20.jpg", + "0000652_21.jpg", + "0000652_22.jpg", + "0000652_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A streetfighter character that acts like Blanka \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001366", + "images": [ + "1001366_00.jpg", + "1001366_01.jpg", + "1001366_02.jpg", + "1001366_03.jpg", + "1001366_04.jpg", + "1001366_05.jpg", + "1001366_06.jpg", + "1001366_07.jpg", + "1001366_08.jpg", + "1001366_09.jpg", + "1001366_10.jpg", + "1001366_11.jpg", + "1001366_12.jpg", + "1001366_13.jpg", + "1001366_14.jpg", + "1001366_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"logo design for vegitable and fruits selling company with yellow green and red colour combination. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2000022", + "images": [ + "2000022_00.jpg", + "2000022_01.jpg", + "2000022_02.jpg", + "2000022_03.jpg", + "2000022_04.jpg", + "2000022_05.jpg", + "2000022_06.jpg", + "2000022_07.jpg", + "2000022_08.jpg", + "2000022_09.jpg", + "2000022_10.jpg", + "2000022_11.jpg", + "2000022_12.jpg", + "2000022_13.jpg", + "2000022_14.jpg", + "2000022_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"looking out window of bullet train outside Riviera ocean passes by \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "3002352", + "images": [ + "3002352_00.jpg", + "3002352_01.jpg", + "3002352_02.jpg", + "3002352_03.jpg", + "3002352_04.jpg", + "3002352_05.jpg", + "3002352_06.jpg", + "3002352_07.jpg", + "3002352_08.jpg", + "3002352_09.jpg", + "3002352_10.jpg", + "3002352_11.jpg", + "3002352_12.jpg", + "3002352_13.jpg", + "3002352_14.jpg", + "3002352_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"\u201cOh, dear!\u201d said Cinderella. \u201cWhen can I\u2013\u201c The stepmother marched into the room. \u201cWhen can you WHAT?\u201d \u201cWell,\u201d said the girl, \u201cwhen will I have time to make my own dress for the ball?\u201d \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "5000085", + "images": [ + "5000085_00.jpg", + "5000085_01.jpg", + "5000085_02.jpg", + "5000085_03.jpg", + "5000085_04.jpg", + "5000085_05.jpg", + "5000085_06.jpg", + "5000085_07.jpg", + "5000085_08.jpg", + "5000085_09.jpg", + "5000085_10.jpg", + "5000085_11.jpg", + "5000085_12.jpg", + "5000085_13.jpg", + "5000085_14.jpg", + "5000085_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The scene is 1980x1080, in the middle of the frame an open window frame hangs in the air (with a size of 560x560) with curtains developing from the wind, the sea below the frame and clouds above, it is also all visible through the frame, evening time, waves. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "7001953", + "images": [ + "7001953_00.jpg", + "7001953_01.jpg", + "7001953_02.jpg", + "7001953_03.jpg", + "7001953_04.jpg", + "7001953_05.jpg", + "7001953_06.jpg", + "7001953_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"make a video of president Lula playing volleyball on the beach \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 2\n" + } + ] + }, + { + "id": "3002859", + "images": [ + "3002859_00.jpg", + "3002859_01.jpg", + "3002859_02.jpg", + "3002859_03.jpg", + "3002859_04.jpg", + "3002859_05.jpg", + "3002859_06.jpg", + "3002859_07.jpg", + "3002859_08.jpg", + "3002859_09.jpg", + "3002859_10.jpg", + "3002859_11.jpg", + "3002859_12.jpg", + "3002859_13.jpg", + "3002859_14.jpg", + "3002859_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a group of skeleton bodied giant persons seeing the dark lightning sky , cinematic movements, realistic.1440 p resolution, 4k , UHD, Motion.4k \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "3001564", + "images": [ + "3001564_00.jpg", + "3001564_01.jpg", + "3001564_02.jpg", + "3001564_03.jpg", + "3001564_04.jpg", + "3001564_05.jpg", + "3001564_06.jpg", + "3001564_07.jpg", + "3001564_08.jpg", + "3001564_09.jpg", + "3001564_10.jpg", + "3001564_11.jpg", + "3001564_12.jpg", + "3001564_13.jpg", + "3001564_14.jpg", + "3001564_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Close up of hands. Astronaut is making a picking gesture. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "4001019", + "images": [ + "4001019_00.jpg", + "4001019_01.jpg", + "4001019_02.jpg", + "4001019_03.jpg", + "4001019_04.jpg", + "4001019_05.jpg", + "4001019_06.jpg", + "4001019_07.jpg", + "4001019_08.jpg", + "4001019_09.jpg", + "4001019_10.jpg", + "4001019_11.jpg", + "4001019_12.jpg", + "4001019_13.jpg", + "4001019_14.jpg", + "4001019_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"drone view of sea, and zoom in to underwater. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001927", + "images": [ + "7001927_00.jpg", + "7001927_01.jpg", + "7001927_02.jpg", + "7001927_03.jpg", + "7001927_04.jpg", + "7001927_05.jpg", + "7001927_06.jpg", + "7001927_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"an anarchist group watching a bank explose in background \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "0001488", + "images": [ + "0001488_00.jpg", + "0001488_01.jpg", + "0001488_02.jpg", + "0001488_03.jpg", + "0001488_04.jpg", + "0001488_05.jpg", + "0001488_06.jpg", + "0001488_07.jpg", + "0001488_08.jpg", + "0001488_09.jpg", + "0001488_10.jpg", + "0001488_11.jpg", + "0001488_12.jpg", + "0001488_13.jpg", + "0001488_14.jpg", + "0001488_15.jpg", + "0001488_16.jpg", + "0001488_17.jpg", + "0001488_18.jpg", + "0001488_19.jpg", + "0001488_20.jpg", + "0001488_21.jpg", + "0001488_22.jpg", + "0001488_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A haunting scene in a forest, a women running away from a lion, shot in black and white 1.33:1 silent aperture \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "1001625", + "images": [ + "1001625_00.jpg", + "1001625_01.jpg", + "1001625_02.jpg", + "1001625_03.jpg", + "1001625_04.jpg", + "1001625_05.jpg", + "1001625_06.jpg", + "1001625_07.jpg", + "1001625_08.jpg", + "1001625_09.jpg", + "1001625_10.jpg", + "1001625_11.jpg", + "1001625_12.jpg", + "1001625_13.jpg", + "1001625_14.jpg", + "1001625_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a pretty lady, quantum electrodynamics, a pretty lady \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "4000254", + "images": [ + "4000254_00.jpg", + "4000254_01.jpg", + "4000254_02.jpg", + "4000254_03.jpg", + "4000254_04.jpg", + "4000254_05.jpg", + "4000254_06.jpg", + "4000254_07.jpg", + "4000254_08.jpg", + "4000254_09.jpg", + "4000254_10.jpg", + "4000254_11.jpg", + "4000254_12.jpg", + "4000254_13.jpg", + "4000254_14.jpg", + "4000254_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Complete black designed Apartment Bedroom with big Windows, sight at a Skyline at night and Rain outside \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "8000510", + "images": [ + "8000510_00.jpg", + "8000510_01.jpg", + "8000510_02.jpg", + "8000510_03.jpg", + "8000510_04.jpg", + "8000510_05.jpg", + "8000510_06.jpg", + "8000510_07.jpg", + "8000510_08.jpg", + "8000510_09.jpg", + "8000510_10.jpg", + "8000510_11.jpg", + "8000510_12.jpg", + "8000510_13.jpg", + "8000510_14.jpg", + "8000510_15.jpg", + "8000510_16.jpg", + "8000510_17.jpg", + "8000510_18.jpg", + "8000510_19.jpg", + "8000510_20.jpg", + "8000510_21.jpg", + "8000510_22.jpg", + "8000510_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A soldier leaping into the fray. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "1001913", + "images": [ + "1001913_00.jpg", + "1001913_01.jpg", + "1001913_02.jpg", + "1001913_03.jpg", + "1001913_04.jpg", + "1001913_05.jpg", + "1001913_06.jpg", + "1001913_07.jpg", + "1001913_08.jpg", + "1001913_09.jpg", + "1001913_10.jpg", + "1001913_11.jpg", + "1001913_12.jpg", + "1001913_13.jpg", + "1001913_14.jpg", + "1001913_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"ice cream inspired by Olympic games \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "5000283", + "images": [ + "5000283_00.jpg", + "5000283_01.jpg", + "5000283_02.jpg", + "5000283_03.jpg", + "5000283_04.jpg", + "5000283_05.jpg", + "5000283_06.jpg", + "5000283_07.jpg", + "5000283_08.jpg", + "5000283_09.jpg", + "5000283_10.jpg", + "5000283_11.jpg", + "5000283_12.jpg", + "5000283_13.jpg", + "5000283_14.jpg", + "5000283_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"young girl is washing clothes in transparent washing machine, we see only water and clothes floating in the air \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002852", + "images": [ + "2002852_00.jpg", + "2002852_01.jpg", + "2002852_02.jpg", + "2002852_03.jpg", + "2002852_04.jpg", + "2002852_05.jpg", + "2002852_06.jpg", + "2002852_07.jpg", + "2002852_08.jpg", + "2002852_09.jpg", + "2002852_10.jpg", + "2002852_11.jpg", + "2002852_12.jpg", + "2002852_13.jpg", + "2002852_14.jpg", + "2002852_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a car crashing into a house cinematic 8k UHD Hyper realistic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "5000365", + "images": [ + "5000365_00.jpg", + "5000365_01.jpg", + "5000365_02.jpg", + "5000365_03.jpg", + "5000365_04.jpg", + "5000365_05.jpg", + "5000365_06.jpg", + "5000365_07.jpg", + "5000365_08.jpg", + "5000365_09.jpg", + "5000365_10.jpg", + "5000365_11.jpg", + "5000365_12.jpg", + "5000365_13.jpg", + "5000365_14.jpg", + "5000365_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"silhouette of conflict between soldiers, cinematic look, dark \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000381", + "images": [ + "7000381_00.jpg", + "7000381_01.jpg", + "7000381_02.jpg", + "7000381_03.jpg", + "7000381_04.jpg", + "7000381_05.jpg", + "7000381_06.jpg", + "7000381_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Little snowflake, falling from the sky \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2002255", + "images": [ + "2002255_00.jpg", + "2002255_01.jpg", + "2002255_02.jpg", + "2002255_03.jpg", + "2002255_04.jpg", + "2002255_05.jpg", + "2002255_06.jpg", + "2002255_07.jpg", + "2002255_08.jpg", + "2002255_09.jpg", + "2002255_10.jpg", + "2002255_11.jpg", + "2002255_12.jpg", + "2002255_13.jpg", + "2002255_14.jpg", + "2002255_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"girl walking on a treadmill and writing on the treadmill desk with plants all around her. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "3001837", + "images": [ + "3001837_00.jpg", + "3001837_01.jpg", + "3001837_02.jpg", + "3001837_03.jpg", + "3001837_04.jpg", + "3001837_05.jpg", + "3001837_06.jpg", + "3001837_07.jpg", + "3001837_08.jpg", + "3001837_09.jpg", + "3001837_10.jpg", + "3001837_11.jpg", + "3001837_12.jpg", + "3001837_13.jpg", + "3001837_14.jpg", + "3001837_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"close up of a male conductor on a bus, cinematic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "7000890", + "images": [ + "7000890_00.jpg", + "7000890_01.jpg", + "7000890_02.jpg", + "7000890_03.jpg", + "7000890_04.jpg", + "7000890_05.jpg", + "7000890_06.jpg", + "7000890_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"video about wishing happy birthday, dancing \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001690", + "images": [ + "7001690_00.jpg", + "7001690_01.jpg", + "7001690_02.jpg", + "7001690_03.jpg", + "7001690_04.jpg", + "7001690_05.jpg", + "7001690_06.jpg", + "7001690_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"1 minute video little mouse play drums \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "3001812", + "images": [ + "3001812_00.jpg", + "3001812_01.jpg", + "3001812_02.jpg", + "3001812_03.jpg", + "3001812_04.jpg", + "3001812_05.jpg", + "3001812_06.jpg", + "3001812_07.jpg", + "3001812_08.jpg", + "3001812_09.jpg", + "3001812_10.jpg", + "3001812_11.jpg", + "3001812_12.jpg", + "3001812_13.jpg", + "3001812_14.jpg", + "3001812_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a super cute man is talking \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002294", + "images": [ + "3002294_00.jpg", + "3002294_01.jpg", + "3002294_02.jpg", + "3002294_03.jpg", + "3002294_04.jpg", + "3002294_05.jpg", + "3002294_06.jpg", + "3002294_07.jpg", + "3002294_08.jpg", + "3002294_09.jpg", + "3002294_10.jpg", + "3002294_11.jpg", + "3002294_12.jpg", + "3002294_13.jpg", + "3002294_14.jpg", + "3002294_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a ghost hunter video, green vision, a forest with a dinosaurs ghost or spirit from past, 8k, camera zoom out, \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "2000558", + "images": [ + "2000558_00.jpg", + "2000558_01.jpg", + "2000558_02.jpg", + "2000558_03.jpg", + "2000558_04.jpg", + "2000558_05.jpg", + "2000558_06.jpg", + "2000558_07.jpg", + "2000558_08.jpg", + "2000558_09.jpg", + "2000558_10.jpg", + "2000558_11.jpg", + "2000558_12.jpg", + "2000558_13.jpg", + "2000558_14.jpg", + "2000558_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"generate five wolves in the dark which stand with frontside to me. Because of darkness, only their eyes to see and the moonshine shows their bodies but in shadow \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001916", + "images": [ + "1001916_00.jpg", + "1001916_01.jpg", + "1001916_02.jpg", + "1001916_03.jpg", + "1001916_04.jpg", + "1001916_05.jpg", + "1001916_06.jpg", + "1001916_07.jpg", + "1001916_08.jpg", + "1001916_09.jpg", + "1001916_10.jpg", + "1001916_11.jpg", + "1001916_12.jpg", + "1001916_13.jpg", + "1001916_14.jpg", + "1001916_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"create a noble hebrew man praying in his tent .4k realistic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2002525", + "images": [ + "2002525_00.jpg", + "2002525_01.jpg", + "2002525_02.jpg", + "2002525_03.jpg", + "2002525_04.jpg", + "2002525_05.jpg", + "2002525_06.jpg", + "2002525_07.jpg", + "2002525_08.jpg", + "2002525_09.jpg", + "2002525_10.jpg", + "2002525_11.jpg", + "2002525_12.jpg", + "2002525_13.jpg", + "2002525_14.jpg", + "2002525_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"The dynamic moment of two civet cats wearing traditional Ewenki costumes riding reindeer, with focused and firm eyes. This is a traditional national reindeer racing competition. The background is the autumn and winter season of Daxinganling Mountains. The painting style is very delicate and vivid and lifelike. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "4001039", + "images": [ + "4001039_00.jpg", + "4001039_01.jpg", + "4001039_02.jpg", + "4001039_03.jpg", + "4001039_04.jpg", + "4001039_05.jpg", + "4001039_06.jpg", + "4001039_07.jpg", + "4001039_08.jpg", + "4001039_09.jpg", + "4001039_10.jpg", + "4001039_11.jpg", + "4001039_12.jpg", + "4001039_13.jpg", + "4001039_14.jpg", + "4001039_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"robot woman wearing a bikinni exploding \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "0000203", + "images": [ + "0000203_00.jpg", + "0000203_01.jpg", + "0000203_02.jpg", + "0000203_03.jpg", + "0000203_04.jpg", + "0000203_05.jpg", + "0000203_06.jpg", + "0000203_07.jpg", + "0000203_08.jpg", + "0000203_09.jpg", + "0000203_10.jpg", + "0000203_11.jpg", + "0000203_12.jpg", + "0000203_13.jpg", + "0000203_14.jpg", + "0000203_15.jpg", + "0000203_16.jpg", + "0000203_17.jpg", + "0000203_18.jpg", + "0000203_19.jpg", + "0000203_20.jpg", + "0000203_21.jpg", + "0000203_22.jpg", + "0000203_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Microorganism shot next to a mushroom, science fiction 8K \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "0000511", + "images": [ + "0000511_00.jpg", + "0000511_01.jpg", + "0000511_02.jpg", + "0000511_03.jpg", + "0000511_04.jpg", + "0000511_05.jpg", + "0000511_06.jpg", + "0000511_07.jpg", + "0000511_08.jpg", + "0000511_09.jpg", + "0000511_10.jpg", + "0000511_11.jpg", + "0000511_12.jpg", + "0000511_13.jpg", + "0000511_14.jpg", + "0000511_15.jpg", + "0000511_16.jpg", + "0000511_17.jpg", + "0000511_18.jpg", + "0000511_19.jpg", + "0000511_20.jpg", + "0000511_21.jpg", + "0000511_22.jpg", + "0000511_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"attractive waitress pouring an endless cup of coffee in a 50s diner \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "2000225", + "images": [ + "2000225_00.jpg", + "2000225_01.jpg", + "2000225_02.jpg", + "2000225_03.jpg", + "2000225_04.jpg", + "2000225_05.jpg", + "2000225_06.jpg", + "2000225_07.jpg", + "2000225_08.jpg", + "2000225_09.jpg", + "2000225_10.jpg", + "2000225_11.jpg", + "2000225_12.jpg", + "2000225_13.jpg", + "2000225_14.jpg", + "2000225_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"supermarket terror dream surreal nightmare benadryl monster \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "5000070", + "images": [ + "5000070_00.jpg", + "5000070_01.jpg", + "5000070_02.jpg", + "5000070_03.jpg", + "5000070_04.jpg", + "5000070_05.jpg", + "5000070_06.jpg", + "5000070_07.jpg", + "5000070_08.jpg", + "5000070_09.jpg", + "5000070_10.jpg", + "5000070_11.jpg", + "5000070_12.jpg", + "5000070_13.jpg", + "5000070_14.jpg", + "5000070_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"snowy flakes in the sky at night \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "2001186", + "images": [ + "2001186_00.jpg", + "2001186_01.jpg", + "2001186_02.jpg", + "2001186_03.jpg", + "2001186_04.jpg", + "2001186_05.jpg", + "2001186_06.jpg", + "2001186_07.jpg", + "2001186_08.jpg", + "2001186_09.jpg", + "2001186_10.jpg", + "2001186_11.jpg", + "2001186_12.jpg", + "2001186_13.jpg", + "2001186_14.jpg", + "2001186_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a colourfull alien in alien world cute \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "6001115", + "images": [ + "6001115_00.jpg", + "6001115_01.jpg", + "6001115_02.jpg", + "6001115_03.jpg", + "6001115_04.jpg", + "6001115_05.jpg", + "6001115_06.jpg", + "6001115_07.jpg", + "6001115_08.jpg", + "6001115_09.jpg", + "6001115_10.jpg", + "6001115_11.jpg", + "6001115_12.jpg", + "6001115_13.jpg", + "6001115_14.jpg", + "6001115_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A superhero woman lands in the jungle with a powerful thud. She looks around and sees the lush greenery surrounding her. She takes a deep breath and prepares for her next move. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "8000764", + "images": [ + "8000764_00.jpg", + "8000764_01.jpg", + "8000764_02.jpg", + "8000764_03.jpg", + "8000764_04.jpg", + "8000764_05.jpg", + "8000764_06.jpg", + "8000764_07.jpg", + "8000764_08.jpg", + "8000764_09.jpg", + "8000764_10.jpg", + "8000764_11.jpg", + "8000764_12.jpg", + "8000764_13.jpg", + "8000764_14.jpg", + "8000764_15.jpg", + "8000764_16.jpg", + "8000764_17.jpg", + "8000764_18.jpg", + "8000764_19.jpg", + "8000764_20.jpg", + "8000764_21.jpg", + "8000764_22.jpg", + "8000764_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"fields of plants high in potasium in van gogh painting style 16:9 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "2001984", + "images": [ + "2001984_00.jpg", + "2001984_01.jpg", + "2001984_02.jpg", + "2001984_03.jpg", + "2001984_04.jpg", + "2001984_05.jpg", + "2001984_06.jpg", + "2001984_07.jpg", + "2001984_08.jpg", + "2001984_09.jpg", + "2001984_10.jpg", + "2001984_11.jpg", + "2001984_12.jpg", + "2001984_13.jpg", + "2001984_14.jpg", + "2001984_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A Animation that shows this text: Welcome To Lofi, Loop If you want to hear more Lofi \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "1001395", + "images": [ + "1001395_00.jpg", + "1001395_01.jpg", + "1001395_02.jpg", + "1001395_03.jpg", + "1001395_04.jpg", + "1001395_05.jpg", + "1001395_06.jpg", + "1001395_07.jpg", + "1001395_08.jpg", + "1001395_09.jpg", + "1001395_10.jpg", + "1001395_11.jpg", + "1001395_12.jpg", + "1001395_13.jpg", + "1001395_14.jpg", + "1001395_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Indiana Jones and the whatever the new thing \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "3002821", + "images": [ + "3002821_00.jpg", + "3002821_01.jpg", + "3002821_02.jpg", + "3002821_03.jpg", + "3002821_04.jpg", + "3002821_05.jpg", + "3002821_06.jpg", + "3002821_07.jpg", + "3002821_08.jpg", + "3002821_09.jpg", + "3002821_10.jpg", + "3002821_11.jpg", + "3002821_12.jpg", + "3002821_13.jpg", + "3002821_14.jpg", + "3002821_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a boy, taking a piece of scrap metal out of a christmas stocking, looking disapointed. Realism style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "0000805", + "images": [ + "0000805_00.jpg", + "0000805_01.jpg", + "0000805_02.jpg", + "0000805_03.jpg", + "0000805_04.jpg", + "0000805_05.jpg", + "0000805_06.jpg", + "0000805_07.jpg", + "0000805_08.jpg", + "0000805_09.jpg", + "0000805_10.jpg", + "0000805_11.jpg", + "0000805_12.jpg", + "0000805_13.jpg", + "0000805_14.jpg", + "0000805_15.jpg", + "0000805_16.jpg", + "0000805_17.jpg", + "0000805_18.jpg", + "0000805_19.jpg", + "0000805_20.jpg", + "0000805_21.jpg", + "0000805_22.jpg", + "0000805_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Trace of red liquid on the ground. Fleeting appearance of a ghostly figure. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 2\n" + } + ] + }, + { + "id": "3001975", + "images": [ + "3001975_00.jpg", + "3001975_01.jpg", + "3001975_02.jpg", + "3001975_03.jpg", + "3001975_04.jpg", + "3001975_05.jpg", + "3001975_06.jpg", + "3001975_07.jpg", + "3001975_08.jpg", + "3001975_09.jpg", + "3001975_10.jpg", + "3001975_11.jpg", + "3001975_12.jpg", + "3001975_13.jpg", + "3001975_14.jpg", + "3001975_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \":black and white picture post nuclear landscape \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 1\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001885", + "images": [ + "7001885_00.jpg", + "7001885_01.jpg", + "7001885_02.jpg", + "7001885_03.jpg", + "7001885_04.jpg", + "7001885_05.jpg", + "7001885_06.jpg", + "7001885_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"the wheels on the bus moving \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "4001144", + "images": [ + "4001144_00.jpg", + "4001144_01.jpg", + "4001144_02.jpg", + "4001144_03.jpg", + "4001144_04.jpg", + "4001144_05.jpg", + "4001144_06.jpg", + "4001144_07.jpg", + "4001144_08.jpg", + "4001144_09.jpg", + "4001144_10.jpg", + "4001144_11.jpg", + "4001144_12.jpg", + "4001144_13.jpg", + "4001144_14.jpg", + "4001144_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"fallen angel crying in a field of flames during apocalypse \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "7000214", + "images": [ + "7000214_00.jpg", + "7000214_01.jpg", + "7000214_02.jpg", + "7000214_03.jpg", + "7000214_04.jpg", + "7000214_05.jpg", + "7000214_06.jpg", + "7000214_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a white butterfly flying upwards from hel and onwards towards a blue sky \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "5000139", + "images": [ + "5000139_00.jpg", + "5000139_01.jpg", + "5000139_02.jpg", + "5000139_03.jpg", + "5000139_04.jpg", + "5000139_05.jpg", + "5000139_06.jpg", + "5000139_07.jpg", + "5000139_08.jpg", + "5000139_09.jpg", + "5000139_10.jpg", + "5000139_11.jpg", + "5000139_12.jpg", + "5000139_13.jpg", + "5000139_14.jpg", + "5000139_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"EXPERT SYSTEMS IN ARTIFICIAL INTELLIGENCE RULES \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "1001263", + "images": [ + "1001263_00.jpg", + "1001263_01.jpg", + "1001263_02.jpg", + "1001263_03.jpg", + "1001263_04.jpg", + "1001263_05.jpg", + "1001263_06.jpg", + "1001263_07.jpg", + "1001263_08.jpg", + "1001263_09.jpg", + "1001263_10.jpg", + "1001263_11.jpg", + "1001263_12.jpg", + "1001263_13.jpg", + "1001263_14.jpg", + "1001263_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Sammy and Whiskers in a meadow made of candy, surrounded by giant lollipops and gumdrop flowers. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "0002334", + "images": [ + "0002334_00.jpg", + "0002334_01.jpg", + "0002334_02.jpg", + "0002334_03.jpg", + "0002334_04.jpg", + "0002334_05.jpg", + "0002334_06.jpg", + "0002334_07.jpg", + "0002334_08.jpg", + "0002334_09.jpg", + "0002334_10.jpg", + "0002334_11.jpg", + "0002334_12.jpg", + "0002334_13.jpg", + "0002334_14.jpg", + "0002334_15.jpg", + "0002334_16.jpg", + "0002334_17.jpg", + "0002334_18.jpg", + "0002334_19.jpg", + "0002334_20.jpg", + "0002334_21.jpg", + "0002334_22.jpg", + "0002334_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"A beautiful afro american woman is lighting a cigarette while she standing in the entrance of a modern elevator. She is wearing grey sweatpants and a olive green bomber jacket. The scene is captured as a wide angle full body fashion photo on 35mm film \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "7001068", + "images": [ + "7001068_00.jpg", + "7001068_01.jpg", + "7001068_02.jpg", + "7001068_03.jpg", + "7001068_04.jpg", + "7001068_05.jpg", + "7001068_06.jpg", + "7001068_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a computerscreen with a popup window saying in english that monitor is going to sleep mode, write correct english text, very detailed, good readable \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "0002950", + "images": [ + "0002950_00.jpg", + "0002950_01.jpg", + "0002950_02.jpg", + "0002950_03.jpg", + "0002950_04.jpg", + "0002950_05.jpg", + "0002950_06.jpg", + "0002950_07.jpg", + "0002950_08.jpg", + "0002950_09.jpg", + "0002950_10.jpg", + "0002950_11.jpg", + "0002950_12.jpg", + "0002950_13.jpg", + "0002950_14.jpg", + "0002950_15.jpg", + "0002950_16.jpg", + "0002950_17.jpg", + "0002950_18.jpg", + "0002950_19.jpg", + "0002950_20.jpg", + "0002950_21.jpg", + "0002950_22.jpg", + "0002950_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Two stars shining bright in the London sky, dark blue sky, magical, 70s dark fantasy style, live action, realistic, cinematic, vintage, HD realistic, ULTRA HD, 70s movie style \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "7000070", + "images": [ + "7000070_00.jpg", + "7000070_01.jpg", + "7000070_02.jpg", + "7000070_03.jpg", + "7000070_04.jpg", + "7000070_05.jpg", + "7000070_06.jpg", + "7000070_07.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a wealthy man swimming in money \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "0002369", + "images": [ + "0002369_00.jpg", + "0002369_01.jpg", + "0002369_02.jpg", + "0002369_03.jpg", + "0002369_04.jpg", + "0002369_05.jpg", + "0002369_06.jpg", + "0002369_07.jpg", + "0002369_08.jpg", + "0002369_09.jpg", + "0002369_10.jpg", + "0002369_11.jpg", + "0002369_12.jpg", + "0002369_13.jpg", + "0002369_14.jpg", + "0002369_15.jpg", + "0002369_16.jpg", + "0002369_17.jpg", + "0002369_18.jpg", + "0002369_19.jpg", + "0002369_20.jpg", + "0002369_21.jpg", + "0002369_22.jpg", + "0002369_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"man on a motorbike crashing into a wall at a high speed, explosion, shrapnel \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 2\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 3\n" + } + ] + }, + { + "id": "4000105", + "images": [ + "4000105_00.jpg", + "4000105_01.jpg", + "4000105_02.jpg", + "4000105_03.jpg", + "4000105_04.jpg", + "4000105_05.jpg", + "4000105_06.jpg", + "4000105_07.jpg", + "4000105_08.jpg", + "4000105_09.jpg", + "4000105_10.jpg", + "4000105_11.jpg", + "4000105_12.jpg", + "4000105_13.jpg", + "4000105_14.jpg", + "4000105_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"a funny blue alien creature with spiral eyes and tentacles for arms. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "3002312", + "images": [ + "3002312_00.jpg", + "3002312_01.jpg", + "3002312_02.jpg", + "3002312_03.jpg", + "3002312_04.jpg", + "3002312_05.jpg", + "3002312_06.jpg", + "3002312_07.jpg", + "3002312_08.jpg", + "3002312_09.jpg", + "3002312_10.jpg", + "3002312_11.jpg", + "3002312_12.jpg", + "3002312_13.jpg", + "3002312_14.jpg", + "3002312_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"gorgon dancing for thirty seconds and screaming trying to get past the screen \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 2\n" + } + ] + }, + { + "id": "4000389", + "images": [ + "4000389_00.jpg", + "4000389_01.jpg", + "4000389_02.jpg", + "4000389_03.jpg", + "4000389_04.jpg", + "4000389_05.jpg", + "4000389_06.jpg", + "4000389_07.jpg", + "4000389_08.jpg", + "4000389_09.jpg", + "4000389_10.jpg", + "4000389_11.jpg", + "4000389_12.jpg", + "4000389_13.jpg", + "4000389_14.jpg", + "4000389_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"In the shadows of deceit, where secrets lie \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "6000660", + "images": [ + "6000660_00.jpg", + "6000660_01.jpg", + "6000660_02.jpg", + "6000660_03.jpg", + "6000660_04.jpg", + "6000660_05.jpg", + "6000660_06.jpg", + "6000660_07.jpg", + "6000660_08.jpg", + "6000660_09.jpg", + "6000660_10.jpg", + "6000660_11.jpg", + "6000660_12.jpg", + "6000660_13.jpg", + "6000660_14.jpg", + "6000660_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"entrepreneur in school not paying attention to his teacher \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "3000410", + "images": [ + "3000410_00.jpg", + "3000410_01.jpg", + "3000410_02.jpg", + "3000410_03.jpg", + "3000410_04.jpg", + "3000410_05.jpg", + "3000410_06.jpg", + "3000410_07.jpg", + "3000410_08.jpg", + "3000410_09.jpg", + "3000410_10.jpg", + "3000410_11.jpg", + "3000410_12.jpg", + "3000410_13.jpg", + "3000410_14.jpg", + "3000410_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"there lived a man 10years old John who was seeking an adventure. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 2\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 1\n" + } + ] + }, + { + "id": "0000738", + "images": [ + "0000738_00.jpg", + "0000738_01.jpg", + "0000738_02.jpg", + "0000738_03.jpg", + "0000738_04.jpg", + "0000738_05.jpg", + "0000738_06.jpg", + "0000738_07.jpg", + "0000738_08.jpg", + "0000738_09.jpg", + "0000738_10.jpg", + "0000738_11.jpg", + "0000738_12.jpg", + "0000738_13.jpg", + "0000738_14.jpg", + "0000738_15.jpg", + "0000738_16.jpg", + "0000738_17.jpg", + "0000738_18.jpg", + "0000738_19.jpg", + "0000738_20.jpg", + "0000738_21.jpg", + "0000738_22.jpg", + "0000738_23.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"an animated area where a palace in a jungle is reflecting in a small water pool and a young boy steps into the pool, disturbing the reflection of the palace, fps 20 \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 3\n factual consistency: 2\n" + } + ] + }, + { + "id": "1002039", + "images": [ + "1002039_00.jpg", + "1002039_01.jpg", + "1002039_02.jpg", + "1002039_03.jpg", + "1002039_04.jpg", + "1002039_05.jpg", + "1002039_06.jpg", + "1002039_07.jpg", + "1002039_08.jpg", + "1002039_09.jpg", + "1002039_10.jpg", + "1002039_11.jpg", + "1002039_12.jpg", + "1002039_13.jpg", + "1002039_14.jpg", + "1002039_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"War in Paris in 2025, filmed by Kubrick , in the street, fire, fireworks, portrait, chaos, cinematographic \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001625", + "images": [ + "6001625_00.jpg", + "6001625_01.jpg", + "6001625_02.jpg", + "6001625_03.jpg", + "6001625_04.jpg", + "6001625_05.jpg", + "6001625_06.jpg", + "6001625_07.jpg", + "6001625_08.jpg", + "6001625_09.jpg", + "6001625_10.jpg", + "6001625_11.jpg", + "6001625_12.jpg", + "6001625_13.jpg", + "6001625_14.jpg", + "6001625_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"at night 2:00am a boy a name virat who crossing the river and also he enjoys the changing seasons amidst the slow waves of the river. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 2\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "6001276", + "images": [ + "6001276_00.jpg", + "6001276_01.jpg", + "6001276_02.jpg", + "6001276_03.jpg", + "6001276_04.jpg", + "6001276_05.jpg", + "6001276_06.jpg", + "6001276_07.jpg", + "6001276_08.jpg", + "6001276_09.jpg", + "6001276_10.jpg", + "6001276_11.jpg", + "6001276_12.jpg", + "6001276_13.jpg", + "6001276_14.jpg", + "6001276_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"ants colony dancing in the rain \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 3\n factual consistency: 1\n" + } + ] + }, + { + "id": "4000441", + "images": [ + "4000441_00.jpg", + "4000441_01.jpg", + "4000441_02.jpg", + "4000441_03.jpg", + "4000441_04.jpg", + "4000441_05.jpg", + "4000441_06.jpg", + "4000441_07.jpg", + "4000441_08.jpg", + "4000441_09.jpg", + "4000441_10.jpg", + "4000441_11.jpg", + "4000441_12.jpg", + "4000441_13.jpg", + "4000441_14.jpg", + "4000441_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"Generate a visually appealing video scene portraying the sentiment of some Canadians getting tired of constant comparisons to their neighbors down south. Use bright scenes to convey the individuality and resilience of the Canadian identity, creating a visually rich narrative without relying on text. \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 2\n text-to-video alignment: 2\n factual consistency: 3\n" + } + ] + }, + { + "id": "2001772", + "images": [ + "2001772_00.jpg", + "2001772_01.jpg", + "2001772_02.jpg", + "2001772_03.jpg", + "2001772_04.jpg", + "2001772_05.jpg", + "2001772_06.jpg", + "2001772_07.jpg", + "2001772_08.jpg", + "2001772_09.jpg", + "2001772_10.jpg", + "2001772_11.jpg", + "2001772_12.jpg", + "2001772_13.jpg", + "2001772_14.jpg", + "2001772_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"video logo with the JRB letter s \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 3\n temporal consistency: 3\n dynamic degree: 3\n text-to-video alignment: 1\n factual consistency: 3\n" + } + ] + }, + { + "id": "1002289", + "images": [ + "1002289_00.jpg", + "1002289_01.jpg", + "1002289_02.jpg", + "1002289_03.jpg", + "1002289_04.jpg", + "1002289_05.jpg", + "1002289_06.jpg", + "1002289_07.jpg", + "1002289_08.jpg", + "1002289_09.jpg", + "1002289_10.jpg", + "1002289_11.jpg", + "1002289_12.jpg", + "1002289_13.jpg", + "1002289_14.jpg", + "1002289_15.jpg" + ], + "conversations": [ + { + "from": "human", + "value": "\nSuppose you are an expert in judging and evaluating the quality of AI-generated videos, \nplease watch the following frames of a given video and see the text prompt for generating the video, \nthen give scores from 5 different dimensions:\n(1) visual quality: the quality of the video in terms of clearness, resolution, brightness, and color\n(2) temporal consistency, the consistency of objects or humans in video\n(3) dynamic degree, the degree of dynamic changes\n(4) text-to-video alignment, the alignment between the text prompt and the video content\n(5) factual consistency, the consistency of the video content with the common-sense and factual knowledge\n\nFor each dimension, output a number from [1,2,3,4], \nin which '1' means 'Bad', '2' means 'Average', '3' means 'Good', \n'4' means 'Real' or 'Perfect' (the video is like a real video)\nHere is an output example:\nvisual quality: 4\ntemporal consistency: 4\ndynamic degree: 3\ntext-to-video alignment: 1\nfactual consistency: 2\n\nFor this video, the text prompt is \"young Latina crying proudly looking ahead, she is in a luxurious living room \",\nall the frames of video are as follows: \n\n \n" + }, + { + "from": "gpt", + "value": "visual quality: 1\n temporal consistency: 1\n dynamic degree: 3\n text-to-video alignment: 2\n factual consistency: 1\n" + } + ] + } +] \ No newline at end of file