As artificial intelligence (AI) image generators become more sophisticated, the challenge of verifying the origin and copyright of the images they produce has intensified. Traditional post-processing watermarks, which are added after an image is created, can be easily removed or bypassed, leaving content creators and platforms vulnerable to misuse. In response, a research team has developed Latent Seal, a novel watermarking framework that embeds watermarks directly into the generation process of latent diffusion models (LDMs), providing a more durable and integrated solution for copyright protection.
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 detailed in a study published in Machine Intelligence Research on June 17, 2026 (DOI: 10.1007/s11633-025-1620-y). Latent Seal works by using a latent-space encoder to blend a full-color RGB watermark into the model's internal representation during the image generation process. A paired decoder then recovers the watermark from protected images, allowing for both generative-content detection and copyright verification.
The team built Latent Seal around Stable Diffusion 2.1, training it on a dataset of 74,247 generated images and their latent representations, sourced from prompts in DiffusionDB and JourneyDB. The system freezes the original denoising network, clones and fine-tunes the variational autoencoder (VAE) decoder, and inserts the watermark encoder into an intermediate decoding block. The decoder is trained to recover the watermark from protected images and return a blank output for unprotected ones, reducing false positives.
In benchmark tests, Latent Seal demonstrated impressive 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 degradation. Recovered watermarks reached 39.19 decibels, 0.9971 structural similarity, and 0.9992 normalized cross-correlation, showing high fidelity in watermark recovery. The method also proved resilient to a wide range of common distortions, including brightness, contrast, and saturation changes, blur, noise, compression, flips, cropping, and rotation. Additionally, Latent Seal added only 7.33 milliseconds during embedding and 2.26 milliseconds during extraction, making it practical for real-time applications.
One of the key advantages of Latent Seal is its ability to carry a full-color image as a watermark, offering more identifying capacity than simple binary signatures. This allows for more detailed provenance information, such as the creator's identity or the model provider's logo. The method also showed consistent performance across different models, including Stable Diffusion XL and Stable Diffusion 3.5, and across various image resolutions.
The researchers emphasize that Latent Seal is designed to make provenance protection an integral part of image creation rather than an afterthought. "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 noted. "Our results suggest that strong watermark recovery and low visual impact can be achieved together."
The potential applications of Latent Seal are broad. It could support provenance checks for commercial image generators, aid in social-media investigations, help resolve copyright disputes, assist in content moderation, and facilitate digital-asset management. The ability to embed a full-color watermark could also serve as a deterrent against unauthorized use, as the watermark can be extracted and compared with the provider's reference to confirm ownership.
However, the current system has limitations. Each new watermark requires retraining, and recovery accuracy slightly decreases as watermark complexity increases. To address this, the researchers propose future work on frequency-domain feature fusion and a lightweight adapter to support arbitrary watermarks without full retraining. They also suggest that Latent Seal is best used alongside other content-authentication tools, such as disclosure policies and metadata standards, rather than as a standalone guarantee.
This research represents a significant step toward more traceable and accountable generative-image systems. By embedding watermarks directly into the generation process, Latent Seal offers a practical and robust solution to the growing challenge of AI content provenance.
