The proliferation of artificial intelligence-generated content (AIGC) has raised critical concerns about authorship, copyright, and misinformation. Traditional post-processing watermarks, applied after image creation, are often vulnerable to removal or manipulation. To address this, a research team has developed Latent Seal, a novel watermarking framework that embeds watermarks during the generation process of latent diffusion models (LDMs), ensuring robustness and minimal visual impact.
Latent Seal operates by integrating a latent-space encoder that blends an RGB watermark into the model's internal representation during image generation. A paired decoder then recovers the watermark from protected images, enabling both generative-content detection and copyright verification. This approach contrasts with conventional methods that attach watermarks post-generation, offering a deeper integration that is more resistant to tampering.
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 the framework around Stable Diffusion 2.1, training on 69,247 generated images and testing on 5,000. They 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 was trained to recover the watermark from protected images and return a blank output for unprotected ones, reducing false positives.
During training, the system simulated ten common distortions, including brightness, contrast, saturation changes, blur, noise, compression, flips, cropping, and rotation. Benchmark tests showed that watermarked images achieved a peak signal-to-noise ratio (PSNR) of 44.29 dB and structural similarity (SSIM) of 0.9933, while recovered watermarks achieved 39.19 dB PSNR, 0.9971 SSIM, and 0.9992 normalized cross-correlation (NCC). Latent Seal maintained the strongest extraction quality across all tested attacks, with minimal computational overhead—adding only 7.33 milliseconds during embedding and 2.26 milliseconds during extraction. Tests on Stable Diffusion XL and 3.5 confirmed consistent performance across models and resolutions.
The authors emphasize that Latent Seal integrates provenance protection into the creation process rather than as 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 stated. "Our results suggest that strong watermark recovery and low visual impact can be achieved together. The next step is to improve recovery for visually complex watermarks and make the framework adaptable to new watermark designs without retraining the full system each time."
The implications of this work are significant for commercial image generators, social media platforms, copyright enforcement, and digital asset management. By embedding high-capacity watermarks that survive common edits, Latent Seal offers a practical solution for tracing AI-generated content. 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 overcome these limitations. They also note that the method works best when combined with disclosure policies and other authentication tools, rather than as a standalone guarantee.
