Research map / Image generation / Autoregressive and latent generation
MSAT-LDM: Toward Transferable High-Fidelity Watermarking for Latent Diffusion Model via Modular Self-Augmented Training
Lu Zhang, Liang Zeng
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 watermark module on free-generation (empty prompt) images instead of external LAION images (ablation) | MSAT-LDM · few-shot transfer to GuoFeng | FID: 6.87 -> 4.88 | Improved |
| train watermark module on free-generation images instead of LAION images (ablation) | MSAT-LDM · few-shot transfer to GuoFeng+ShuiMo LoRA | FID: 13.2 -> 8.30 | Improved |
| train watermark module on free-generation images instead of LAION images (ablation) | MSAT-LDM · few-shot transfer to Realistic | PSNR: 37.2 -> 36.8 | No clear effect |
| transfer pretrained watermark module instead of re-initializing it before few-shot fine-tuning (ablation) | MSAT-LDM · few-shot transfer to GuoFeng | FID: 150 -> 4.88 | Improved |
| transfer pretrained watermark module instead of re-initializing it before few-shot fine-tuning (ablation) | MSAT-LDM · few-shot transfer to GuoFeng | adversarial bit accuracy: 0.99 -> 0.99 | No clear effect |
| MSAT-LDM vs FSW under same few-shot fine-tuning | FSW baseline · few-shot transfer to GuoFeng | FID: 169 -> 4.88 | Improved |
| MSAT-LDM vs Stable Signature on few-shot transfer | Stable Signature baseline · few-shot transfer to GuoFeng | FID: 11.7 -> 4.88 | Improved |
| MSAT-LDM vs prior diffusion-native watermarks for image quality | SD v1.5 · AI-generated prompts | FID: 9.49 -> 2.35 | Improved |
| train on LDM-generated data (free generation) vs real LAION images (ablation) | watermark module training · AI-generated test prompts | FID and adversarial bit accuracy | Improved |
| free generation vs LAION-prompt generated data as training distribution (ablation) | watermark module training · AI-generated test prompts | robustness at small sample sizes | Mixed |
| training-free Gaussian Shading vs MSAT-LDM on geometric attack | Gaussian Shading · perspective warp attack | bit accuracy: 0.67 vs 0.98 | Hurt |
Closest papers by meaning
- StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models (NeurIPS 2025 · similarity 0.89)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.89)
- WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.89)
- Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models (CVPR 2025 · similarity 0.88)
- AquaLoRA: Toward White-box Protection for Customized Stable Diffusion Models via Watermark LoRA (ICML 2024 · similarity 0.88)
- Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models (CVPR 2025 · similarity 0.88)
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.
