A collaborative research team has introduced Latent Seal, a novel watermarking framework that integrates robust watermarks directly into the generation process of latent diffusion models (LDMs), rather than applying them post hoc. The technique, detailed in a study published in Machine Intelligence Research, is designed to enhance copyright protection and traceability of AI-generated imagery without compromising visual quality.
The proliferation of generative image systems has made it increasingly difficult to verify the origin and authorship of digital content. Traditional post-processing watermarks are easy to implement but remain vulnerable to removal or bypass. In-generation methods integrate protection more deeply, yet many suffer from limited information capacity or degrade under common image manipulations. Latent Seal addresses these challenges by embedding a high-capacity RGB watermark into the model's internal latent space during image creation, while a paired decoder extracts the mark only from protected images.
The framework, developed by researchers from Macao Polytechnic University, Guangdong University of Technology, Jinan University, and the Institute of Automation, Chinese Academy of Sciences, was built around Stable Diffusion 2.1. The team assembled a dataset of 74,247 generated images and their latent representations, using prompts from DiffusionDB and JourneyDB. The system freezes the original denoising network, clones and fine-tunes the variational autoencoder decoder, and inserts a latent-space watermark encoder into an intermediate decoding block. A separate decoder is trained to recover the target watermark from protected images and return a blank output for unprotected ones, reducing false positives.
In benchmark tests, Latent Seal demonstrated exceptional performance. Watermarked images achieved a peak signal-to-noise ratio of 44.29 decibels and a structural similarity index of 0.9933, indicating minimal visual distortion. Recovered watermarks achieved 39.19 decibels, 0.9971 structural similarity, and 0.9992 normalized cross-correlation. The method retained the strongest extraction quality across all tested attacks—including brightness, contrast, saturation changes, blur, noise, compression, flips, cropping, and rotation—while adding only 7.33 milliseconds during embedding and 2.26 milliseconds during extraction. Further tests on Stable Diffusion XL and Stable Diffusion 3.5 confirmed consistent performance across models and resolutions.
The authors emphasize that Latent Seal is designed to make provenance protection intrinsic to image creation. "The aim is to preserve the visual quality users expect while giving model providers a practical way to verify origin after images have been edited or shared," they stated. "Our results suggest that strong watermark recovery and low visual impact can be achieved together."
The potential impact of this technology is significant for commercial image generators, social media platforms, copyright enforcement, content moderation, and digital asset management. By embedding a full-color image watermark, Latent Seal offers greater identifying capacity than simple binary signatures and shows resilience to routine edits, helping marks survive ordinary online sharing. However, the current system requires retraining for each new watermark, and recovery accuracy diminishes with more complex watermark textures. The researchers propose future enhancements, including frequency-domain feature fusion and a lightweight adapter for arbitrary watermarks, to address these limitations.
In practice, Latent Seal would be most effective when combined with disclosure policies, metadata standards, and other content-authentication tools, rather than as a standalone guarantee. The study, published with DOI 10.1007/s11633-025-1620-y, received funding from the Science and Technology Development Fund of Macau SAR and Macao Polytechnic University. The research appears in Machine Intelligence Research, a journal sponsored by the Institute of Automation, Chinese Academy of Sciences.
