RPG-RT: Red-Teaming Text-to-Image Systems by Rule-based Preference Modeling

This repository contains the code for Red-Teaming Text-to-Image Systems by Rule-based Preference Modeling, accepted by NeurIPS 2025.

Installation

We provide the environment.yaml to create a conda environment. You could create the environment with:

conda env create -f environment.yaml

Fine-tuning for RPG-RT

You can choose your target T2I systems in scripts/run_model.sh and run RPG-RT with:

bash scripts/run_model.sh

Evaluation

Evaluation code for RPG-RT could be found in the notebook evaluation/evaluation.ipynb, including the ASR, CLIP similarity (CS), FID and per​​p​​le​​x​​ity (PPL).

Citation

Please feel free to email us at caoyichuan@amss.ac.cn. If you find this useful, please consider citing our work!

@article{cao2025red,
  title={Red-Teaming Text-to-Image Systems by Rule-based Preference Modeling},
  author={Cao, Yichuan and Miao, Yibo and Gao, Xiao-Shan and Dong, Yinpeng},
  journal={arXiv preprint arXiv:2505.21074},
  year={2025}
}