RAVEN: erasing invisible watermarks from AI-generated images via novel view synthesis

Researchers propose the RAVEN framework, which reformulates invisible watermark removal as a novel view synthesis problem, achieving watermark suppression superior to 14 baseline attacks across 15 watermarking schemes while maintaining higher perceptual quality. The idea is to generate perceptually consistent views of the same semantic content from shifted viewpoints, similar to re-observing a scene from different angles, thereby naturally removing embedded watermarks while preserving visual fidelity. The method is a zero-shot diffusion framework based on a frozen pretrained model, applying controlled geometric transformations in latent space and using view-guided correspondence attention to maintain structural consistency during reconstruction, requiring no access to the detector or knowledge of the watermark at runtime. The authors note that watermarks robust to pixel-space and frequency-domain attacks remain vulnerable under semantics-preserving viewpoint transformations.