Research map / Image generation / Autoregressive and latent generation
Guidance Watermarking for Diffusion Models
Enoal Gesny, Eva Giboulot, Teddy Furon, Vivien Chappelier
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 |
|---|---|---|---|
| add watermark guidance during diffusion (G-VS) | VideoSeal decoder on SD2 · SD2 | decoding capacity (bits): +37.7 bits (to 212.2) | Improved |
| add guidance (G-SSig) | Stable Signature decoder on SD2 · SD2 | -log10(PFA) @ PD=0.9: +12.2 (to 16.3) | Improved |
| add watermark guidance | SD2 / Flux / Sana · image quality | FID and CLIP: SD2 FID 5.0 -> 2.2; Flux 9.5 -> 9.3; Sana 4.3 -> 4.1 | No clear effect |
| guide with G-VS instead of Gaussian Shading | SD2 · Rotation 90 (multi-bit) | capacity (bits): 0 -> 194 | Improved |
| guide with G-VS instead of Gaussian Shading | SD2 · Identity (multi-bit) | capacity (bits): 221 -> 222 | No clear effect |
| guide with G-VS instead of Tree-Rings | SD2 · Crop 50% (zero-bit) | -log10(PFA) @ PD=0.9: 0.4 -> 101.9 | Improved |
| restrict guidance to last 15 steps (ablation) | G-VS on Sana · Sana | -log10(PFA) @ PD=0.9: 96.4 -> 70.0 | Hurt |
| restrict guidance to last 5 steps (ablation) | G-VS on Sana · Sana | -log10(PFA) @ PD=0.9: 96.4 -> 6.2 | Hurt |
| use full backpropagation through diffusion instead of simplified identity gradient (ablation) | G-VS on Sana · Sana | PSNR / detection: PSNR 23.5 -> 8.8; -log10(PFA) 96.4 -> 49.1 | Hurt |
| include unknown attack (rotation 90, median filter) in guidance augmentation set | G-SSig · SD2 and Sana zero-bit | detectability | Improved |
Closest papers by meaning
- WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.91)
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.90)
- ROBIN: Robust and Invisible Watermarks for Diffusion Models with Adversarial Optimization (NeurIPS 2024 · similarity 0.89)
- AquaLoRA: Toward White-box Protection for Customized Stable Diffusion Models via Watermark LoRA (ICML 2024 · similarity 0.89)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Secure Seed-Based Multi-bit Watermarking for Diffusion Models from First Principles (NeurIPS 2026 · 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.
