T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models

The code requires Python 3.10.13 and the packages listed in requirements.txt. Install them with:

pip install -r requirements.txt

Note: To run T2SMark on Stable Diffusion v3.5 Medium, upgrade diffusers from 0.21.4 to 0.32.0. Otherwise, StableDiffusion3Pipeline will not be supported.

python run.py --name test

python run_sd35.py --name test_sd35

The --name argument is required. For additional options, see option.py.

Our code includes built-in evaluation of TPR, bit accuracy, and CLIP score. For LPIPS and FID metrics, we recommend using the following repositories:

• LPIPS score: https://github.com/richzhang/PerceptualSimilarity.git
• FID: https://github.com/mseitzer/pytorch-fid.git

Note: The COCO prompts and ground-truth images used in our experiments are available here, sourced from the Tree-Ring Watermark repository.

The code for all baseline methods is provided below:

• DwtDct, DwtDctSvd, RivaGan: https://github.com/ShieldMnt/invisible-watermark.git
• Stable Signature: https://github.com/facebookresearch/stable_signature.git
• Tree-Ring Watermark: https://github.com/YuxinWenRick/tree-ring-watermark.git
• Gaussian Shading: https://github.com/bsmhmmlf/Gaussian-Shading.git
• PRC-Watermark: https://github.com/XuandongZhao/PRC-Watermark.git

T2SMark focuses on watermark design, while FARI focuses on fast and robust watermark recovery. If you are interested in inversion-based verification for diffusion watermarking, check out FARI: Robust One-Step Inversion for Watermarking in Diffusion Models.

We borrow the code from Tree-Ring Watermark and Gaussian Shading. We appreciate the authors for sharing their code.

If our work assists your research, feel free to give us a star ⭐ or cite us using:

@inproceedings{yang2025t2smark,
  author = {Yang, Jindong and Fang, Han and Zhang, Weiming and Yu, Nenghai and Chen, Kejiang},
  booktitle = {Advances in Neural Information Processing Systems},
  title = {T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models},
  pages = {118642--118668},
  volume = {38},
  year = {2025}
}