Learning to Watermark in the Latent Space of Generative Models

Abstract

Existing approaches for watermarking AI-generated outputs often rely on post-hoc methods applied in pixel space, introducing computational overhead and potential visual artifacts. In this work, we explore latent space watermarking and its distillation into generative models. We introduce DistSeal, a unified approach for latent watermarking that works across both diffusion and autoregressive models. We demonstrate that post-hoc watermarking methods can be applied directly in the latent domain and that these latent watermarkers can be effectively distilled intothe generative pipeline (either into the generative model itself or into the latent decoder) enabling in-model watermarking. The resulting post-hoc latent method achieves robustness comparable to pixel-space baselines while offering better imperceptibility and a 20× speedup. Our experiments further reveal that distilling latent watermarkers outperforms distilling pixel-space ones, providing a solution that is both more efficient and more robust. Code and model will be made available at https://github.com/facebookresearch/distseal