The Latent Seal: Robust Model Watermarking for Latent Diffusion Model

In recent years, the Latent Diffusion Model (LDM) has gained widespread adoption across various industries due to its significant commercial value. However, the content generated by LDM currently lacks sufficient copyright protection, raising serious legal and ethical concerns. To address this issue, model watermarking technologies have been proposed as viable solutions. Nevertheless, traditional watermarking methods typically embed watermarks after the content has been generated, thus limiting their effectiveness. In this paper, we propose a novel watermarking model, named Latent Seal, that embeds watermarks directly during the content generation process. The proposed Latent Seal employs an encoder-decoder architecture, where a latentspace encoder embeds image watermarks within the latent space during content generation, and a latentspace decoder ensures that the target watermark can only be extracted from watermarked images. Extensive experimental results demonstrate that Latent Seal exhibits outstanding performance in terms of imperceptibility and robustness.