ash12321/seedream-4.5-specialist-h100

SeeDream 4.5 Specialist 🎯

Specialist detector for SeeDream 4.5 (unfrozen last layer, 97-99% accuracy)

H100-Optimized Specialist Detector with Unfrozen Last Layer

Architecture

• Base: Vision Transformer (ViT-base-patch16-224)
• Layers 0-10: FROZEN (pretrained features)
• Layer 11: UNFROZEN (learns generator-specific patterns)
• Classifier: UNFROZEN (768 → 2)
• Trainable Parameters: ~100,000
• Training Data: 400 images (320 train / 40 val / 40 test)

Performance

Metric	Score
Test Accuracy	1.0000 (100.00%)
Precision	1.0000
Recall	1.0000
F1 Score	1.0000

Features

✅ Specialist detector (trained only on SeeDream) ✅ Unfrozen last attention layer for generator-specific features ✅ Test-Time Augmentation (TTA) enabled ✅ Heavy regularization (prevents overfitting on small dataset) ✅ H100-optimized training

Usage

from transformers import ViTForImageClassification, ViTImageProcessor
from PIL import Image
import torch

model = ViTForImageClassification.from_pretrained("ash12321/seedream-4.5-specialist-h100")
processor = ViTImageProcessor.from_pretrained("google/vit-base-patch16-224")

image = Image.open("image.jpg")
inputs = processor(images=image, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.softmax(outputs.logits, dim=1)
    
if probs[0][1] > 0.5:
    print(f"AI-Generated (SeeDream): {probs[0][1]:.2%}")
else:
    print(f"Real: {probs[0][0]:.2%}")

Training Details

• Training Time: 2.1 minutes
• Best Epoch: 16
• Device: H100 GPU
• Unique data split (seed=456)

Limitations

⚠️ This is a specialist detector trained ONLY for SeeDream.

Does NOT detect:

• Other AI image generators
• General synthetic images

For comprehensive AI detection, use as part of an ensemble with other specialist detectors.