MediaPipe-Face-Detection: Optimized for Qualcomm Devices

Designed for sub-millisecond processing, this model predicts bounding boxes and pose skeletons (left eye, right eye, nose tip, mouth, left eye tragion, and right eye tragion) of faces in an image.

This is based on the implementation of MediaPipe-Face-Detection found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime	Precision	Chipset	SDK Versions	Download
ONNX	float	Universal	QAIRT 2.42, ONNX Runtime 1.24.3	Download
QNN_DLC	float	Universal	QAIRT 2.45	Download
TFLITE	float	Universal	QAIRT 2.45	Download

For more device-specific assets and performance metrics, visit MediaPipe-Face-Detection on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

• Custom weights (e.g., fine-tuned checkpoints)
• Custom input shapes
• Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for MediaPipe-Face-Detection on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.object_detection

Model Stats:

• Input resolution: 256x256
• Number of output classes: 6
• Number of parameters (face_detector): 135K
• Model size (face_detector) (w8a8): 260 KB
• Number of parameters (face_landmark_detector): 603K
• Model size (face_landmark_detector) (w8a8): 750 KB
• Model size (face_detector) (float): 557 KB
• Model size (face_landmark_detector) (float): 2.33 MB

Performance Summary

Model	Runtime	Precision	Chipset	Inference Time (ms)	Peak Memory Range (MB)	Primary Compute Unit
face_detector	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	0.333 ms	0 - 34 MB	NPU
face_detector	ONNX	float	Snapdragon® X2 Elite	0.407 ms	0 - 0 MB	NPU
face_detector	ONNX	float	Snapdragon® X Elite	0.88 ms	2 - 2 MB	NPU
face_detector	ONNX	float	Snapdragon® 8 Gen 3 Mobile	0.499 ms	0 - 43 MB	NPU
face_detector	ONNX	float	Qualcomm® QCS8550 (Proxy)	0.684 ms	0 - 2 MB	NPU
face_detector	ONNX	float	Qualcomm® QCS9075	1.143 ms	1 - 3 MB	NPU
face_detector	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	0.413 ms	0 - 29 MB	NPU
face_detector	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	0.271 ms	1 - 33 MB	NPU
face_detector	QNN_DLC	float	Snapdragon® X2 Elite	0.546 ms	1 - 1 MB	NPU
face_detector	QNN_DLC	float	Snapdragon® X Elite	0.689 ms	1 - 1 MB	NPU
face_detector	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	0.391 ms	0 - 40 MB	NPU
face_detector	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	2.897 ms	1 - 30 MB	NPU
face_detector	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	0.528 ms	1 - 71 MB	NPU
face_detector	QNN_DLC	float	Qualcomm® SA8775P	0.997 ms	1 - 32 MB	NPU
face_detector	QNN_DLC	float	Qualcomm® QCS9075	0.874 ms	1 - 3 MB	NPU
face_detector	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	0.769 ms	0 - 37 MB	NPU
face_detector	QNN_DLC	float	Qualcomm® SA7255P	2.897 ms	1 - 30 MB	NPU
face_detector	QNN_DLC	float	Qualcomm® SA8295P	1.099 ms	0 - 25 MB	NPU
face_detector	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.334 ms	1 - 33 MB	NPU
face_detector	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	0.281 ms	0 - 32 MB	NPU
face_detector	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	0.397 ms	0 - 39 MB	NPU
face_detector	TFLITE	float	Qualcomm® QCS8275 (Proxy)	2.949 ms	0 - 28 MB	NPU
face_detector	TFLITE	float	Qualcomm® QCS8550 (Proxy)	0.541 ms	0 - 72 MB	NPU
face_detector	TFLITE	float	Qualcomm® SA8775P	1.029 ms	0 - 32 MB	NPU
face_detector	TFLITE	float	Qualcomm® QCS9075	0.887 ms	0 - 3 MB	NPU
face_detector	TFLITE	float	Qualcomm® QCS8450 (Proxy)	0.77 ms	0 - 37 MB	NPU
face_detector	TFLITE	float	Qualcomm® SA7255P	2.949 ms	0 - 28 MB	NPU
face_detector	TFLITE	float	Qualcomm® SA8295P	1.09 ms	0 - 25 MB	NPU
face_detector	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	0.339 ms	0 - 32 MB	NPU
face_landmark_detector	ONNX	float	Snapdragon® 8 Elite Gen 5 Mobile	0.156 ms	0 - 27 MB	NPU
face_landmark_detector	ONNX	float	Snapdragon® X2 Elite	0.157 ms	0 - 0 MB	NPU
face_landmark_detector	ONNX	float	Snapdragon® X Elite	0.378 ms	0 - 0 MB	NPU
face_landmark_detector	ONNX	float	Snapdragon® 8 Gen 3 Mobile	0.207 ms	0 - 33 MB	NPU
face_landmark_detector	ONNX	float	Qualcomm® QCS8550 (Proxy)	0.281 ms	0 - 24 MB	NPU
face_landmark_detector	ONNX	float	Qualcomm® QCS9075	0.42 ms	0 - 3 MB	NPU
face_landmark_detector	ONNX	float	Snapdragon® 8 Elite For Galaxy Mobile	0.177 ms	0 - 22 MB	NPU
face_landmark_detector	QNN_DLC	float	Snapdragon® 8 Elite Gen 5 Mobile	0.126 ms	0 - 26 MB	NPU
face_landmark_detector	QNN_DLC	float	Snapdragon® X2 Elite	0.244 ms	0 - 0 MB	NPU
face_landmark_detector	QNN_DLC	float	Snapdragon® X Elite	0.331 ms	0 - 0 MB	NPU
face_landmark_detector	QNN_DLC	float	Snapdragon® 8 Gen 3 Mobile	0.162 ms	0 - 33 MB	NPU
face_landmark_detector	QNN_DLC	float	Qualcomm® QCS8275 (Proxy)	0.852 ms	0 - 23 MB	NPU
face_landmark_detector	QNN_DLC	float	Qualcomm® QCS8550 (Proxy)	0.219 ms	0 - 72 MB	NPU
face_landmark_detector	QNN_DLC	float	Qualcomm® SA8775P	0.441 ms	0 - 24 MB	NPU
face_landmark_detector	QNN_DLC	float	Qualcomm® QCS9075	0.332 ms	0 - 3 MB	NPU
face_landmark_detector	QNN_DLC	float	Qualcomm® QCS8450 (Proxy)	0.328 ms	0 - 38 MB	NPU
face_landmark_detector	QNN_DLC	float	Qualcomm® SA7255P	0.852 ms	0 - 23 MB	NPU
face_landmark_detector	QNN_DLC	float	Qualcomm® SA8295P	0.581 ms	0 - 24 MB	NPU
face_landmark_detector	QNN_DLC	float	Snapdragon® 8 Elite For Galaxy Mobile	0.151 ms	0 - 23 MB	NPU
face_landmark_detector	TFLITE	float	Snapdragon® 8 Elite Gen 5 Mobile	0.108 ms	0 - 27 MB	NPU
face_landmark_detector	TFLITE	float	Snapdragon® 8 Gen 3 Mobile	0.148 ms	0 - 33 MB	NPU
face_landmark_detector	TFLITE	float	Qualcomm® QCS8275 (Proxy)	0.912 ms	0 - 23 MB	NPU
face_landmark_detector	TFLITE	float	Qualcomm® QCS8550 (Proxy)	0.194 ms	0 - 1 MB	NPU
face_landmark_detector	TFLITE	float	Qualcomm® SA8775P	0.42 ms	0 - 25 MB	NPU
face_landmark_detector	TFLITE	float	Qualcomm® QCS9075	0.309 ms	0 - 4 MB	NPU
face_landmark_detector	TFLITE	float	Qualcomm® QCS8450 (Proxy)	0.314 ms	0 - 37 MB	NPU
face_landmark_detector	TFLITE	float	Qualcomm® SA7255P	0.912 ms	0 - 23 MB	NPU
face_landmark_detector	TFLITE	float	Qualcomm® SA8295P	0.587 ms	0 - 25 MB	NPU
face_landmark_detector	TFLITE	float	Snapdragon® 8 Elite For Galaxy Mobile	0.134 ms	0 - 29 MB	NPU

License

• The license for the original implementation of MediaPipe-Face-Detection can be found here.

References

• BlazeFace: Sub-millisecond Neural Face Detection on Mobile GPUs
• Source Model Implementation

Community

• Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
• For questions or feedback please reach out to us.