DTR: Dynamic Tree-Ring Watermarking Framework for Diffusion-Based Video Generation

The growing capabilities of diffusion-based text-to-video models have raised significant concerns about copyright protection and the traceability of synthetic video content. To address these concerns, existing watermarking techniques have been developed to invisibly embed information within video content. However, many of these techniques lack resilience when faced with different distortions. This makes existing methods unsuitable for real-world applications, where videos frequently undergo compression, scaling, and other transformations. In response to this challenge, we propose DTR, a training-free Dynamic Tree-Ring watermarking framework designed to enhance the robustness of video watermarks while preserving video quality. Our DTR embeds the predefined user-specific key into the Fourier domain of the initial latent variables. Then, the invertible Fourier transformation ensures the invisible nature of such an insertion. To further improve video quality, the key is divided and embedded as discrete segments across different initial latents, thereby minimizing distribution shifts caused by the watermark insertion. Comprehensive experiments demonstrate the superiority of the DTR framework compared to traditional approaches in terms of robustness, especially under the H.264 and H.265 compression standards, while maintaining high video quality and inference efficiency.