k-aisi-anti-deepfake/dear-lsun-inpaint

DEAR Diagnostic Inpaint Set (dear-lsun-inpaint)

Diagnostic data for the dissection step of DEAR ("Dissect and Prune: Enhancing Robustness in AI-Generated Image Detection", ICML 2026). Each real LSUN image has a random rectangular region inpainted with Stable Diffusion 1.5, so real and generated pixels coexist in one image under a known mask. DEAR uses these paired images and masks to measure per-channel Regional Activation Discrepancy (RAD).

The image folders are shipped as tar archives (Hugging Face allows at most 10000 files per folder).

• —Paper: https://arxiv.org/abs/2606.10309
• —Code: https://github.com/dahyedahye/dear
• —Full assembly and usage: see docs/DATASET.md in the code repo.

Contents

lsun_inpaint_sd.tar     # inpainted images, -> data/train/fake/lsun_inpaint_sd/
lsun_inpaint_mask.tar   # binary masks,     -> data/train/fake/lsun_inpaint_mask/

Files are paired by filename stem ({stem}_inpaint.png and {stem}_mask.png).

Usage

huggingface-cli download k-aisi-anti-deepfake/dear-lsun-inpaint \
    --repo-type dataset --local-dir ./dear-lsun-inpaint

mkdir -p data/train/fake
tar xf ./dear-lsun-inpaint/lsun_inpaint_sd.tar   -C data/train/fake/
tar xf ./dear-lsun-inpaint/lsun_inpaint_mask.tar -C data/train/fake/

You can also regenerate the set with scripts/inpaint_data_gen/ in the code repo.

Source and license

Released under CC BY-NC 4.0 for research and non-commercial use. The base real images are from LSUN. The inpainted content is generated by Stable Diffusion 1.5, so these images additionally carry the use-based restrictions of the CreativeML OpenRAIL-M license. See NOTICE and LICENSE for the full terms.

Citation

@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}
}