As artificial intelligence (AI) image generators become capable of producing realistic visuals at scale, verifying the origin and authorship of these images has become a pressing concern. Traditional post-hoc watermarking methods, which add marks after an image is created, are vulnerable to removal or bypass. Now, a research team has introduced Latent Seal, a watermarking framework that embeds watermarks directly into the generation process of latent diffusion models (LDMs), offering a more durable solution for copyright protection and content provenance.
Latent Seal operates by inserting a latent-space encoder that blends a red-green-blue (RGB) watermark into the model's internal representation during image generation. A paired decoder then recovers the mark from protected images, enabling both detection of AI-generated content and copyright verification. The approach is designed to be visually unobtrusive, preserving image quality while making the watermark resilient to common edits and distortions.
The research, published in Machine Intelligence Research on June 17, 2026, was conducted by scientists from Macao Polytechnic University, Guangdong University of Technology, Jinan University, and the Institute of Automation, Chinese Academy of Sciences. The team built Latent Seal around Stable Diffusion 2.1, using a dataset of 74,247 generated images and their latent representations. 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.
To test robustness, the researchers simulated ten common distortions during training, including brightness changes, blur, noise, compression, flips, cropping, and rotation. In benchmark tests, 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, demonstrating high fidelity even after attacks.
Latent Seal also showed strong performance across different models, including Stable Diffusion XL and Stable Diffusion 3.5, and added only 7.33 milliseconds during embedding and 2.26 milliseconds during extraction, making it practical for real-time applications.
The development of Latent Seal has significant implications for the AI industry. As generative models become more accessible, the ability to trace content back to its source is crucial for copyright enforcement, content moderation, and digital asset management. By embedding watermarks during generation, providers can verify origin even after images have been edited or shared online.
However, the current system requires retraining for each new watermark, and recovery accuracy decreases with more complex watermark designs. The researchers propose future improvements, including frequency-domain feature fusion and lightweight adapters for arbitrary watermarks. They also emphasize that Latent Seal should be used alongside other content-authentication tools, such as metadata standards and disclosure policies, to provide a comprehensive solution.
This research was funded by the Science and Technology Development Fund of Macau SAR (No. 0053/2025/RIB2) and Macao Polytechnic University (No. RP/FCA-04/2024). The full study is available at https://doi.org/10.1007/s11633-025-1620-y.
