As artificial intelligence-generated images become increasingly realistic and widespread, the question of authorship and copyright has grown more pressing. A new watermarking framework called Latent Seal aims to address this by embedding robust watermarks directly into the image generation process of latent diffusion models (LDMs), rather than attaching them as a post-processing step. The approach, developed by researchers from Macao Polytechnic University, Guangdong University of Technology, Jinan University, and the Institute of Automation, Chinese Academy of Sciences, is detailed in a study published in Machine Intelligence Research on June 17, 2026 (DOI: 10.1007/s11633-025-1620-y).
Traditional watermarking methods, which add marks after an image is created, are relatively simple to implement but can be removed or bypassed. In-generation techniques offer stronger protection but often suffer from limited information capacity or fragility under common edits such as compression, cropping, or rotation. Latent Seal overcomes these challenges by integrating a latent-space encoder that blends an RGB watermark into the model’s internal representation during generation, paired with a decoder that recovers the mark only from protected images. This allows for both generative-content detection and copyright verification without visibly degrading image quality.
In testing, watermarked images achieved a peak signal-to-noise ratio (PSNR) of 44.29 decibels and a structural similarity index (SSIM) of 0.9933, indicating minimal visual impact. The recovered watermarks demonstrated high fidelity with a PSNR of 39.19 decibels, SSIM of 0.9971, and normalized cross-correlation of 0.9992. The system also maintained the strongest extraction quality across all simulated attacks, which included brightness, contrast, saturation changes, blur, noise, compression, flips, cropping, and rotation. Latent Seal added only 7.33 milliseconds during embedding and 2.26 milliseconds during extraction, making it practical for real-time applications.
Built around Stable Diffusion 2.1, the researchers trained the system on 74,247 generated images, with 69,247 used for training and 5,000 for testing. The framework froze the original denoising network, cloned and fine-tuned the variational autoencoder (VAE) decoder, and inserted the watermark encoder into an intermediate decoding block. A separate decoder learned to recover the watermark from protected images and return a blank output for unprotected ones, reducing false detections. The method also showed consistent performance across other models like Stable Diffusion XL and Stable Diffusion 3.5, demonstrating its versatility across different architectures and image resolutions.
The researchers emphasize that Latent Seal is designed to make provenance protection an integral part of 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 said. “Our results suggest that strong watermark recovery and low visual impact can be achieved together.” They also acknowledge limitations, noting that the current system requires retraining for each new watermark and that recovery accuracy decreases with more complex watermark designs. Future work will explore frequency-domain feature fusion and lightweight adapters to handle arbitrary watermarks without full retraining.
The implications of this research are significant for commercial image generators, social media platforms, copyright enforcement, and digital asset management. By embedding watermarks during generation, providers can trace the origin of images even after they have been edited or shared online. Latent Seal’s ability to carry a full-color image offers greater identifying capacity than simple binary signatures, and its resilience to common manipulations could help marks survive ordinary online distribution. However, the researchers stress that the method is best used in combination with other content-authentication tools, such as disclosure policies and metadata standards, rather than as a standalone guarantee.
As generative AI continues to evolve, tools like Latent Seal represent a critical step toward ensuring accountability and trust in digital content. By integrating watermarking into the generation process, this approach could become a cornerstone of copyright protection in the age of AI.
