k-aisi-anti-deepfake/dear-checkpoints

DEAR Checkpoints

Model weights for DEAR ("Dissect and Prune: Enhancing Robustness in AI-Generated Image Detection", ICML 2026). DEAR is a post-hoc feature-selection method that dissects a pretrained detector's channels by Regional Activation Discrepancy, bilaterally prunes the extremes, and refines the linear classifier.

• Paper: https://arxiv.org/abs/2606.10309
• Code: https://github.com/dahyedahye/dear

Files

corvi/model_best.pth     # base Corvi detector (ResNet-50)
rajan/model_best.pth     # base Rajan detector (ResNet-50, AlignedForensics)
dear_c/model_best.pth    # DEAR-c (ours, Corvi base)
dear_r/model_best.pth    # DEAR-r (ours, Rajan base)

Usage

huggingface-cli download k-aisi-anti-deepfake/dear-checkpoints --local-dir checkpoints
python scripts/inference.py --model dear_c --input path/to/image.png

See the code repo for the full inference and evaluation pipeline.

Training data

The base detectors and DEAR models were trained on LSUN and MS COCO real images, LDM fake images from the AlignedForensics training data (originally the DMimageDetection latent diffusion set by GRIP-UNINA, Corvi et al.), the AlignedForensics aligned pairs (Rajan), and the SD 1.5 inpainted diagnostic set (DEAR refinement). The backbone is a ResNet-50 from GRIP-UNINA initialized from torchvision ImageNet weights.

Source and license

Released under CC BY-NC 4.0 for research and non-commercial use. Because the DEAR models are refined on Stable Diffusion 1.5 generated data, they also carry the use-based restrictions of the CreativeML OpenRAIL-M license. See NOTICE and LICENSE for the full attribution and terms.

Citation

For the DEAR models (dear_c, dear_r), cite the DEAR paper.

@inproceedings{kim2026dissect,
  title     = {Dissect and Prune: Enhancing Robustness in AI-Generated Image Detection},
  author    = {Kim, Dahye and Choi, Jaehyun and Seong, Hyun Seok and Kim, Seongho and Lee, Donghun and Yi, Sungwon and Choi, Jang-Ho},
  booktitle = {Proceedings of the Forty-third International Conference on Machine Learning},
  year      = {2026},
  url       = {https://arxiv.org/abs/2606.10309}
}

For the base detectors, cite the original works. The Corvi detector (corvi) and its training data are from Corvi et al. The Rajan detector (rajan) is from Sundara Rajan et al. (AlignedForensics).

@inproceedings{corvi2023detection,
  title     = {On the detection of synthetic images generated by diffusion models},
  author    = {Corvi, Riccardo and Cozzolino, Davide and Zingarini, Giada and Poggi, Giovanni and Nagano, Koki and Verdoliva, Luisa},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year      = {2023}
}

@inproceedings{rajan2025aligned,
  title     = {Aligned Datasets Improve Detection of Latent Diffusion-Generated Images},
  author    = {Sundara Rajan, Anirudh and Ojha, Utkarsh and Schloesser, Jedidiah and Lee, Yong Jae},
  booktitle = {International Conference on Learning Representations (ICLR)},
  year      = {2025}
}