WMVLM: Evaluating Diffusion Model Image Watermarks Based on Vision-Language Models

The researchers propose WMVLM, which they describe as the first unified interpretable evaluation framework for diffusion model image watermarks based on vision-language models (VLMs). The framework evaluates residual watermarks by artifact strength and anti-erasure capability, evaluates semantic watermarks by latent space distribution shift, and progressively achieves classification, scoring, and interpretable text generation through three-stage training. Experiments show that WMVLM surpasses state-of-the-art VLMs and generalizes strongly across datasets, diffusion models, and watermarking methods.