ReImage: Robust Image Self-Recovery against Tampering using Watermark Generation with Pixel Shuffling

Official pretrained weights for ReImage, a neural watermarking-based image self-recovery framework that embeds a shuffled version of the target image into itself as a watermark, enabling tamper localization and recovery of the original content.

• 📄 Paper: arXiv:2511.22936
• 🌐 Project page: https://eurominyoung186.github.io/ReImage/
• 💻 Code: https://github.com/EuroMinyoung186/ReImage

Files

File	Description
models/200000_G.pth	Network weights at iteration 200k (use this for inference)
training_state/200000.state	Optimizer/scheduler state at iteration 200k (only needed to resume training)

Download

pip install -U huggingface_hub
hf download Eurong2/ReImage --local-dir ../experiments/invertible_pipeline

The repo mirrors the directory layout expected by the test configs (an experiments/ directory next to the code repository root):

../experiments/invertible_pipeline/models/200000_G.pth
../experiments/invertible_pipeline/training_state/200000.state

Then run inference following the instructions in the code repository:

python test2.py -opt options/test/test.yml

Citation

@article{kim2025reimage,
    title = "Robust Image Self-Recovery against Tampering using Watermark Generation with Pixel Shuffling",
    author = "Kim, Minyoung and Seo, Paul Hongsuck",
    journal = "arXiv preprint arXiv:2511.22936",
    year = "2025"
}