k-aisi-anti-deepfake/aigi-detection-ldm

DEAR-c Training Data (aigi-detection-ldm)

Training data for the base Corvi detector used in DEAR ("Dissect and Prune: Enhancing Robustness in AI-Generated Image Detection", ICML 2026).

This repository hosts the LSUN real images and the LDM fake images. The image folders are shipped as tar archives (Hugging Face allows at most 10000 files per folder). The DEAR detectors also use COCO real images, which are not re-hosted here (download COCO 2017 from the official site, see below).

• —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

train_real_lsun.tar        # -> data/train/real/lsun/
val_real_lsun.tar          # -> data/val/real/lsun/
train_fake_ldm.tar         # -> data/train/fake/ldm/
val_fake_ldm.tar           # -> data/val/fake/ldm/
coco_train_filenames.txt   # the COCO 2017 images we used (train subset)
coco_val_filenames.txt     # the COCO 2017 images we used (val subset)

Usage

huggingface-cli download k-aisi-anti-deepfake/aigi-detection-ldm \
    --repo-type dataset --local-dir ./aigi-dl

mkdir -p data/train/real data/val/real data/train/fake data/val/fake
tar xf ./aigi-dl/train_real_lsun.tar -C data/train/real/
tar xf ./aigi-dl/val_real_lsun.tar   -C data/val/real/
tar xf ./aigi-dl/train_fake_ldm.tar  -C data/train/fake/
tar xf ./aigi-dl/val_fake_ldm.tar    -C data/val/fake/

Then add the COCO real images (official COCO 2017) under data/train/real/coco/ and data/val/real/coco/. Use coco_train_filenames.txt and coco_val_filenames.txt to select the same subset we used. See the code repo for details.

Source and license

Released under CC BY-NC 4.0 for research and non-commercial use. The LSUN real images are released for research use. The LDM fake images are built from the AlignedForensics training data (AniSundar18/aligned_forensic_trainingdata), which is based on the latent diffusion training set of DMimageDetection by GRIP-UNINA (Corvi et al., Apache-2.0). When using the LDM fakes, please also cite Corvi et al., "On the detection of synthetic images generated by diffusion models". See NOTICE for the full attribution and per-source terms.

Citation

@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}
}