luvmelo/Face_Classifier_100real_100comfyui

Real vs Synthetic Face Classifier

This package contains tools to classify images as Real or Synthetic (AI-Generated). It specifically targets high-quality generations (e.g., from ComfyUI/Flux) that are often difficult to detect.

Models

There are two trained models available. When uploading, you can choose one or both.

1. trained_classifier_random (Baselne ResNet):

   • Trained on standard dataset.
   • Good for general detection, but might miss high-quality "hard" fakes.

2. trained_classifier_densenet_hard_mining (Recommended):

   • Trained on an expanded dataset including "Hard Negatives" (high-quality fakes that fooled the baseline).
   • Architecture: DenseNet121.
   • Higher robustness.

How to Share/Upload

Option A: Hugging Face (Simplest for Sharing)

1. Create a new Model repository on Hugging Face.
2. Upload the contents of this deployment folder to the repo.
3. Crucial: You must also upload the model_weights.pth and model_info.json from the trained model folder (e.g., trained_classifier_densenet_hard_mining) into the root of your Hugging Face repo (or a subfolder).

Option B: GitHub

1. Create a new repo.
2. Commit this deployment folder.
3. Warning: The .pth weight files are large (~30-50MB). GitHub standard file limit is 100MB, so they should fit. If they are larger, you need Git LFS.
   • git lfs install
   • git lfs track "*.pth"

Usage Guide (For your Users)

1. Installation

pip install -r requirements.txt

2. Run Inference

Use the inference.py script to classify a single image.

Command:

python inference.py path/to/image.jpg --model path/to/model_folder

Example:

# Assuming you downloaded the model weights into a folder named 'my_model'
python inference.py test_image.png --model my_model

3. Output

The script will print the confidence scores for both "Real" and "Synthetic". It automatically handles background removal (using rembg) before classification, matching the training pipeline.