Research map / Image generation / Autoregressive and latent generation
Learning to Watermark in the Latent Space of Generative Models
Sylvestre-Alvise Rebuffi, Tuan A. Tran, Valeriu Lacatusu, Pierre Fernandez, Tomáš Souček, Nikola Jovanović, Tom Sander, Hady Elsahar, Alexandre Mourachko
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 |
|---|---|---|---|
| watermark in latent space instead of pixel space (post-hoc, DCAE) | post-hoc watermarker · DCAE 512x512 | Inference time per image on CPU: 63 ms -> 3 ms (20x speedup) | Improved |
| watermark in latent space instead of pixel space (post-hoc, DCAE) | post-hoc watermarker · DCAE 512x512 | Avg bit accuracy (%): 97.78 -> 95.18 | Hurt |
| watermark in latent space instead of pixel space (post-hoc, DCAE) | post-hoc watermarker · DCAE combined attacks | Bit accuracy (%): 97.29 -> 84.28 | Hurt |
| make embedder input-dependent instead of coverless (ablation) | post-hoc latent watermarker · DCAE 512x512 | Avg bit accuracy (%): 91.81 -> 95.18 | Improved |
| apply latent embedder after quantization instead of before (ablation) | post-hoc latent watermarker · RAR-XL | Avg bit accuracy (%): 90.60 -> 93.96 | Improved |
| distill latent watermarker instead of pixel watermarker into latent decoder (no extractor loss) (ablation) | DCAE latent decoder · DCAE | Avg bit accuracy (%): 51.38 -> 94.41 | Improved |
| add extractor loss (weight 0.1) to latent decoder distillation (ablation) | DCAE latent decoder from latent teacher · DCAE | Avg bit accuracy (%): 94.41 -> 96.77 | Improved |
| add strong extractor loss (weight 1.0) when distilling pixel watermarker into decoder (ablation) | DCAE latent decoder from pixel teacher · DCAE | FID: bit acc 51.38 -> 92.95; FID 10.72 -> 17.08 | Mixed |
| add extractor loss (weight 0.1) to latent decoder distillation (ablation) | RAR-XL latent decoder from latent teacher · RAR-XL | Avg bit accuracy (%): 81.40 -> 90.57 | Improved |
| distill pixel watermarker into RAR latent decoder with strong extractor loss (ablation) | RAR-XL latent decoder from pixel teacher · RAR-XL | Avg bit accuracy (%): 97.27 -> 71.25 | Hurt |
| distill latent watermarker into the generative model by fine-tuning on watermarked latents | DCAE diffusion model · DCAE | Avg bit accuracy (%) / FID: 95.18 -> 94.78 bit acc; FID 11.42 -> 10.90 | Improved |
| distill latent watermarker into the generative model by fine-tuning on watermarked tokens | RAR-XL · RAR-XL | Avg bit accuracy (%) / FID: 90.60 -> 89.13 bit acc; FID 3.56 -> 3.32 | Improved |
Closest papers by meaning
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.88)
- How to Trace Latent Generative Model Generated Images without Artificial Watermark? (ICML 2024 · similarity 0.88)
- A Locally Tokenized Generative Model for Robust Time-Series Watermarking (NeurIPS 2026 · similarity 0.88)
- WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.88)
- RAW: A Robust and Agile Plug-and-Play Watermark Framework for AI-Generated Images with Provable Guarantees (NeurIPS 2024 · 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.
