🔍 Learning to Falsify

Hierarchical Multiple Instance Learning for High Resolution AI-Generated Image Detection
🎉 ACM Multimedia 2026

LTF learns local evidence of image synthesis with a frozen CLIP backbone and two levels of multiple instance learning. This repository includes training, evaluation, original-resolution inference, FATAS aggregation, and the Gradio demo.

📰 News

• ACM Multimedia 2026 — 🎉 LTF has been accepted to ACM Multimedia 2026!
• 2026-11-04 — Code, GenImage checkpoint integration, and the 🤗 interactive demo are available.

🚀 Quick start

Python 3.11 is recommended. Install PyTorch for your CUDA environment first if needed.

pip install -r requirements.txt
python predict.py path/to/image.jpg

This release provides the GenImage checkpoint (historical mean accuracy: 98.39%). It and the CLIP backbone download automatically. Assets are pinned and verified in assets.json. For offline use:

python predict.py image.jpg --weights /path/to/ltf.pth --backbone /path/to/ViT-L-14.pt

The default uses CLIP block 24, a global view and 224×224 local tiles. Full blocks do not overlap; when needed, a final block is aligned to the right/bottom edge. Only those boundary blocks may overlap. level1, level2 and fusion scores are returned; class 1 means generated. See checkpoint and reproduction notes before comparing scores with other benchmarks reported in the paper.

🛠️ Training and evaluation

python train.py --config configs/train.yaml --dr3-root data/reconstructions --genimage-root data/GenImage
# Multi-GPU: torchrun --nproc_per_node=4 train.py --config configs/train.yaml
python evaluate.py --data data/GenImage --output outputs/genimage

Data layout and configuration cover paired reconstruction data, benchmark folders, feature layers and resume options. Evaluation saves per-image scores and per-subset accuracy, AUC, AP, real false-positive rate and fake recall. The summary is an unweighted subset average.

🤗 Demo and FATAS

pip install -r demo/requirements.txt
python demo/app.py
# Existing independent tile scores: image,tile,probability
python tools/fatas.py --input tile_scores.csv --k 3 --output scores.json

The demo includes 8 clickable GenImage validation examples (4 real and 4 generated). Clicking an image runs prediction and attention analysis. It also shows tiles, MIL attention, individual aggregation heads and patch weights. Space setup includes a local packaging command. FATAS averages the top-k generated probabilities from an independent tile scorer. It is separate from LTF fusion and the demo's top-four display.

The project page also includes a Robustness Lab and Meet the Heads: measured image transformations and eight linked L1 attention maps. Data and regeneration instructions.

📝 TODO

• Release checkpoints and evaluation profiles for additional benchmarks.
• Complete full training and benchmark reproduction checks.
• Add data preparation scripts.

📚 Citation

@inproceedings{chen2026ltf,
  title={Learning to Falsify: Hierarchical Multiple Instance Learning for High Resolution AI-Generated Image Detection},
  author={Chen, Haoran and Gao, Daiqi and Li, Xinyuan and Wang, Yongwei and Zhu, Lei and Cai, Rui and Dong, Jianfeng},
  booktitle={Proceedings of the 34th ACM International Conference on Multimedia},
  year={2026},
  doi={10.1145/3767308.3836365},
  eprint={2611.04529},
  archivePrefix={arXiv}
}

📖 Code: MIT. Paper figures: CC BY 4.0. Interface icons and badges: sources.