Edge-Constrained Controllable Stable Diffusion Watermarking

Existing diffusion-based watermarking methods often experience latent distribution shifts, leading to structural distortions in the generated images. To tackle this issue, we propose the Edge-Constrained Controllable Stable Diffusion Watermarking (EC-CSW) framework, which incorporates edge features as geometric priors. This framework integrates multiscale edge features from the original, watermark-free images as geometric constraints, dynamically guiding the denoising process of the diffusion model. This method effectively balances the concealment of the watermark with the visual plausibility of the generated images. Experimental results, validated using 1,000 diverse text-image pairs based on the Stable Diffusion-v2 model, show that the proposed framework significantly mitigates the problems of unreasonable images caused by watermark embedding. This research offers innovative insights into the application of diffusion models in the field of watermarking, achieving a dual optimization of copyright protection and visual quality, and holds significant theoretical and practical implications.