C2P-CLIP: Injecting Category Common Prompt in CLIP to Enhance Generalization in Deepfake Detection

Chuangchuang Tan, Renshuai Tao, Huan Liu, Guanghua Gu, Baoyuan Wu, Yao Zhao, Yunchao Wei

Beijing Jiaotong University, YanShan University, CUHK

⭐ If our code is helpful to you, please help star this repo. Thanks! 🤗

News 🆕

• Pretrained models & Text links & Dataset link

Overall Pipeline

---

🛠️ Installation

1) Main Environment (Training & Detection)

conda create -n c2pclip python=3.10.14 -y
conda activate c2pclip
pip install -r requirements.txt

📂 Data Preparation

• Prepare your dataset (e.g., GenImage, UniversalFakeDetect).
• Download Genimage_CNNDetection_CLIP_prefix_caption.tar.gz from the provided Google Drive link.
• Download CLIP weights (ViT-L/14) from Hugging Face.

---

🚀 Usage

1) Training

Train C2P-CLIP on GenImage and UniversalFakeDetect.

conda activate c2pclip

./train_genimage.sh

./train_UniversalFakeDetect.sh

2) Inference / Testing

conda activate c2pclip

python inference.py \
  --dataroot ./datasets/GenImage/test/ \
  --model_path ./checkpoints/c2p_clip_genimage/last_model.pth

3) Feature Analysis (Decoding & Visualization)

conda activate c2pclip

# Decode features to text
python decode_clipfeature_image.py \
  --image_path ./assets/DALLE/DALLE_2_Cowboy_In_Swamp_Close_Up_Outpaint_1.png \
  --cal_detection_feat

# Visualization (t-SNE)
CUDA_VISIBLE_DEVICES=1 python draw_tsne_kmean.py \
  --draw_data_path ./tsne_png \
  --image_path ./stylegan_tsne_data  \
  --save_name stylegan_test \
  --legend stylegan-bedroom-real stylegan-bedroom-fake stylegan-car-real stylegan-car-fake stylegan-cat-real stylegan-cat-fake \
  --do_extract --do_fit --draw_text 0

📝 Citation

If you find this code or paper helpful, please cite:

@inproceedings{tan2025c2p,
  title={C2p-clip: Injecting category common prompt in clip to enhance generalization in deepfake detection},
  author={Tan, Chuangchuang and Tao, Renshuai and Liu, Huan and Gu, Guanghua and Wu, Baoyuan and Zhao, Yao and Wei, Yunchao},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={7},
  pages={7184--7192},
  year={2025}
}

🙏 Acknowledgments

This repository borrows partially from the CLIPCap, NPR.