Hidden in the Noise: Two-Stage Robust Watermarking for Images
We demonstrate that the initial noise used by the diffusion model to generate images can serve as a distortion-free watermark, but detecting it requires comparing the inversed noise with all previously used noises. To address these limitations, we propose a two-stage framework, WIND, that divides initial noises into groups and embeds group identifiers into them. For detection, we first retrieve the group and then identify the matching noise.
To install the dependencies, run the setup.sh script:
chmod +x setup.sh
./setup.shpython WIND_full.py --onlinepython WIND_fast.py --onlinePlease check the initial_noise branch for the code related to other experiments discussed in the paper.