DGS-Net: Distillation-Guided Gradient Surgery for CLIP Fine-Tuning in AI-Generated Image Detection

Jiazhen Yan¹, Ziqiang Li¹, Fan Wang², Boyu Wang¹, Ziwen He¹, Zhangjie Fu^1‡

^‡Corresponding author

¹School of Computer Science, Nanjing University of Information Science and Technology ²University of Macau

📰 News

• [2026-05-01]🎉🎉🎉 DGS-Net is accepted by ICML 2026 Spotlight

🔥 Todo List

• Inference code
• Pretrained models
• Training code

🚀 Quick Start

1. Installation

conda create -n DGS-Net -y python=3.10

conda activate DGS-Net
pip3 install torch torchvision
pip install -r requirements.txt 

2.Getting datasets

Datasets	Paper	Url
UniversalFakeDetect	Towards Universal Fake Image Detectors that Generalize Across Generative Models (CVPR 2023)	Google Drive
AIGCDetectBench	PatchCraft: Exploring Texture Patch for Efficient AI-generated Image Detection	ModelScope
AIGIBench	Is Artificial Intelligence Generated Image Detection a Solved Problem? (NeurIPS 2025)	Huggingface/Baidu Netdisk

3.Download Weights

Place them under checkpoints/: Google Drive

4.Inference

Of course, you need to change [DetectionTests] in test.py when testing.

We also present our inference results in log_test.log. [Note]: Following re-inference, we observed an approximately 0.1 discrepancy relative to the original result. The file log_test.log contains the results obtained after re-inference.

python test.py --model_path ./checkpoints/model_epoch_step2.pth

⏳ Training

The training set uses ProGAN & SDv1.4 from AIGIBench (Is Artificial Intelligence Generated Image Detection a Solved Problem?, NeurIPS 2025) Huggingface/Baidu Netdisk.

stage 1: python train.py --name  5class-car-cat-chair-horse-sdv1.4 --dataroot /home/HDD/yjz/dataset/AIGIBench/train  --classes  car,cat,chair,horse,sdv1.4 --train_stage 0 --niter 3
stage 2: python train.py --name  5class-car-cat-chair-horse-sdv1.4 --dataroot /home/HDD/yjz/dataset/AIGIBench/train  --classes  car,cat,chair,horse,sdv1.4 --train_stage 1 --niter 1

During the stage 2 of training, it is necessary to change the model_path of stage 1 in models/clip_models: self.clip_model_frozen.load_state_dict(torch.load("/path/checkpoints/model_epoch_stage1.pth"), strict=True)

📚 Citation

@article{yan2025dgs,
  title={DGS-Net: Distillation-Guided Gradient Surgery for CLIP Fine-Tuning in AI-Generated Image Detection},
  author={Yan, Jiazhen and Li, Ziqiang and Wang, Fan and Wang, Boyu and He, Ziwen and Fu, Zhangjie},
  journal={arXiv preprint arXiv:2511.13108},
  year={2025}
}

📬 Contact

If you have any question about this project, please feel free to contact 247918horizon@gmail.com