Research map / Image generation / Autoregressive and latent generation
WMVLM: Evaluating Diffusion Model Image Watermarking via Vision-Language Models
Zijin Yang, Yu Sun, Kejiang Chen, Jiawei Zhao, Jun Jiang, Weiming Zhang, Nenghai Yu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train Qwen3-VL-8B with WMVLM three-stage pipeline instead of using zero-shot VLMs | VLM watermark evaluator · Residual watermarks, SD v2.1 | quality PLCC / SRCC, security accuracy: RobustWide-like zero-shot VINE 0.067/0.052/0.500 (GPT-5) -> 0.983/0.986/1.000 | Improved |
| add explainability cold start (S2) after pre-training (S1) (ablation) | WMVLM · RobustWide / VINE / RingID | PLCC and accuracy: RobustWide PLCC 0.967 -> 0.869 | Hurt |
| add GRPO (S3) after S1+S2 (ablation) | WMVLM · RivaGAN and HiDDeN residual | PLCC / SRCC: RivaGAN PLCC 0.185 -> 0.885 | Improved |
| add GRPO (S3) after S1+S2 (ablation) | WMVLM · Semantic Lossless | quality/security accuracy: 0.844 -> 0.910 | Improved |
| add pre-training stage S1 before cold start and GRPO (ablation) | WMVLM · Semantic watermarks | accuracy: Lossless 0.000/0.000 -> 0.910/0.910 | Improved |
| add cold start S2 before GRPO (instead of GRPO alone) (ablation) | WMVLM · All methods | format reward / accuracy | Improved |
| evaluate on MS-COCO prompts instead of SD prompts (cross-dataset) | WMVLM · RingID semantic | accuracy: 0.852/0.852 -> 0.746/0.746 | Hurt |
| evaluate on SD v1.4 instead of SD v2.1 (cross-model) | WMVLM · Stable Signature residual | PLCC: 0.981 -> 0.368 | Hurt |
| evaluate on unseen watermarking methods (cross-method) | WMVLM · GaussMarker, T2SMark, Ori | accuracy: 0.956/0.956, 0.930/0.930, 0.874/0.874 | No clear effect |
| scale zero-shot open-source VLM size (Qwen3-VL 4B to 32B) | Qwen3-VL · Residual quality scoring | PLCC | No clear effect |
Closest papers by meaning
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.90)
- Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models (CVPR 2025 · similarity 0.90)
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.90)
- WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.89)
- AquaLoRA: Toward White-box Protection for Customized Stable Diffusion Models via Watermark LoRA (ICML 2024 · similarity 0.89)
- Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models (CVPR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
