Are Semantic Watermarks for Diffusion Models Resilient to Layout Control?

Abstract

Semantic watermarking methods embed information into generated images by modifying the initial latent noise, subtly modifying the output images. However, the widespread use of layout control techniques, such as ControlNets, raises questions about the applicability of semantic watermarking with layout control. After all, if semantic watermarks are really realized as meaningful changes in images (such as its layout), external layout specifications (e.g. through edge maps), could destroy the watermark information during denoising. This work empirically evaluates two semantic watermarking approaches—Tree-Ring Watermarking and Gaussian Shading—under various ControlNet-guided generation settings. Our results show that while ControlNets can slightly degrade watermark strength, both watermarking approaches remain largely detectable, demonstrating the potential viability of semantic watermarks even under strong layout constraints.