What is it about?
MaXsive is a training-free generative watermarking technique for diffusion models, designed to have both high capacity and high robustness. We have shown that an X-shaped template can be injected into the image generation process to effectively recover from rotation, scaling, and translation (RST) distortions. This is a new approach that avoids the meticulously repetitive ring patterns used by other methods. We show that MaXsive outperforms existing algorithms on robustness benchmarks.
Featured Image
Photo by Growtika on Unsplash
Why is it important?
The rapid proliferation of AI image generators necessitates effective watermarking to protect copyright and trace distribution. Unfortunately, existing solutions are often inadequate: they either lack robustness against removal and manipulation or they cannot scale to support the vast number of users and unique identifiers required. MaXive directly addresses this gap by providing a scalable, high-capacity, and robust watermarking solution.
Perspectives
Writing this paper was exciting because we were stuck on the same problem as everyone else: current training-free watermarks force a bad trade-off between robustness and capacity. If they could survive a simple image rotation, they couldn't store enough unique IDs for real-world use. The breakthrough came when we found out that we can couple a separate "X-shape" template with the diffusion sampling process. This design allowed us to finally achieve both high capacity and strong robustness, and moreover, preserve the image quality. We hope this provides a truly practical tool that model owners can use to protect their work and trace content at scale.
Po-Yuan Mao
University of Bath
Read the Original
This page is a summary of: MaXsive: High-Capacity and Robust Training-Free Generative Image Watermarking in Diffusion Models, October 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3746027.3755266.
You can read the full text:
Contributors
The following have contributed to this page
