MaxMark: High-Capacity Diffusion-Native Watermarking via Robust and Invertible Latent Embedding - CVPR

Model Description

This repository contains the model weights from our CVPR paper on latent space watermarking. The models demonstrate robust watermarking capabilities in the latent representation space with different bit configurations for embedding imperceptible watermarks with invertible neural networks.

Model Details

• Model Type: INN
• Framework: PyTorch
• Model Description:
  • 16384-bit configuration: latent-4x64x64_margin10.0_wm_16384bits.pth
  • 8192-bit configuration: latent-4x64x64_margin10.0_wm_8192bits.pth

Key Features

• Latent Space Processing: Operates on 4×64×64 latent representations
• Margin Parameter: 10.0 margin for watermark robustness
• Multiple Bit Depths: Supports less than 16384 and 8192-bit watermarks, while 8192-bit model has less impact on latent.
• Robust Watermarking: Imperceptible watermarks resistant to common attacks

Citation

If you use this model in your research, please cite our paper on OpenReview:

@inproceedings{maxmark2026,
  title={MaxMark: High-Capacity Diffusion-Native Watermarking via Robust and Invertible Latent Embedding},
  author={Xuanhang Chang, Zhonghao Yang, Cheng Zhuo, Yu Li},
  booktitle={CVPR 2026},
  year={2026},
  url={https://openreview.net/pdf?id=8yhxO86boG}
}

License

This model is released under the MIT License.