Research map / Image generation / Autoregressive and latent generation
The Stable Signature: Rooting Watermarks in Latent Diffusion Models
Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, Teddy Furon
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 |
|---|---|---|---|
| fine-tune LDM decoder with a message loss and perceptual loss | LDM decoder · MS-COCO text-to-image | FID: -0.3 FID | No clear effect |
| merge watermarking into decoder instead of post-hoc HiDDeN | Text-to-image watermarking · MS-COCO text-to-image | bit accuracy (combined edit): 0.98 -> 0.92 | Hurt |
| merge watermarking into decoder instead of post-hoc Dct-Dwt | Text-to-image watermarking · MS-COCO text-to-image | bit accuracy (combined edit): 0.51 -> 0.92 | Improved |
| increase perceptual loss weight lambda_i from 0.025 to 0.8 (ablation) | Decoder fine-tuning · Combined edit | bit accuracy: 0.95 -> 0.85 | Hurt |
| increase perceptual loss weight lambda_i from 0.025 to 0.8 (ablation) | Decoder fine-tuning · Generated images | PSNR: 24.6 -> 31.4 | Improved |
| add JPEG to the attack simulation layer during extractor pre-training (ablation) | Watermark extractor · JPEG 50 test | bit accuracy: 0.50 -> 0.90 | Improved |
| add rotation to the attack simulation layer during extractor pre-training (ablation) | Watermark extractor · Rotation 90 test | bit accuracy: 0.56 -> 0.99 | Improved |
| add crop to the attack simulation layer during extractor pre-training (ablation) | Watermark extractor · Crop 0.1 test | bit accuracy: 1.00 -> 1.00 | No clear effect |
| apply PCA whitening to extractor outputs (ablation) | Watermark extractor · Vanilla images | i.i.d. Bernoulli bit assumption | Improved |
| attack with the same auto-encoder as the LDM | Watermark removal attack · Image-level attack | watermark erasure at quality: PSNR>29 with watermark erased | Hurt |
| increase number of users N for identification (ablation) | Identification · Identification at FPR 1e-6 | accuracy: tau 41 (N=1) -> 44 (N=1000) | Hurt |
| model purification by fine-tuning to remove watermark | Network-level attack · Model purification | bit accuracy vs image quality | No clear effect |
Closest papers by meaning
- How to Trace Latent Generative Model Generated Images without Artificial Watermark? (ICML 2024 · similarity 0.91)
- Attack-Resilient Image Watermarking Using Stable Diffusion (NeurIPS 2024 · similarity 0.91)
- ROBIN: Robust and Invisible Watermarks for Diffusion Models with Adversarial Optimization (NeurIPS 2024 · similarity 0.90)
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.90)
- 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)
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.
