Research map / Image generation / Autoregressive and latent generation
LaWa: Using Latent Space for In-Generation Image Watermarking
Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, Yong Zhang
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 |
|---|---|---|---|
| embed watermark in latent space of frozen decoder (in-generation) instead of post-generation encoder-decoder or optimization methods | LaWa · AI-generated images, 32-bit | average bit accuracy: 0.92 -> 0.99 vs SSL/RivaGan | Improved |
| in-generation embedding instead of post-generation processing | LaWa · AI-generated images | embedding time (ms): SSL 870 -> LaWa 1 | Improved |
| use LaWa as post-generation instead of in-generation | LaWa · AI-generated images | average bit accuracy: 0.99 -> 0.99 (32-bit); 0.97 -> 0.97 (48-bit) | No clear effect |
| LaWa vs FNNS under 10% random crop at 48 bits | LaWa · AI-generated images, 48-bit | bit accuracy: FNNS 0.96 vs LaWa 0.91 | Hurt |
| increase message length from 32 to 128 bits (ablation) | LaWa · AI-generated images | bit accuracy (combined attack): 0.97 -> 0.87 | Hurt |
| increase quality-robustness weight lambda from 1.0 to 10.0 (ablation) | LaWa · AI-generated images | bit accuracy on combined attack / PSNR: bit acc 0.65 -> 0.99; PSNR 34.76 -> 28.47 | Mixed |
| add Critic loss to MSE and LPIPS losses (ablation) | LaWa · AI-generated images | PSNR/SSIM: 29.47/0.47 -> 33.52/0.86 | Improved |
| add LPIPS loss to MSE plus Critic (ablation) | LaWa · AI-generated images | LPIPS/SIFID: 0.10/0.05 -> 0.04/0.02 | Improved |
| add LPIPS loss to MSE plus Critic (ablation) | LaWa · AI-generated images | bit accuracy on combined attack: 0.99 -> 0.94 | Hurt |
| raise image MSE loss weight lambda_I above 0.1 (ablation) | LaWa · AI-generated images | LPIPS/SIFID and robustness: bit acc comb. 0.94 -> 0.84 at lambda_I=1.0 | Hurt |
| LaWa vs Stable Signature against autoencoder removal attack (KL-VAE f8) | LaWa · learning-based attacks | bit accuracy: 55% -> 85% | Improved |
Later papers that compared with this method
6 later accepted papers report Lawa(48) as a baseline in their results tables; 5 of them report a better number on most of the results they share with it.
- GenPTW: Latent Image Watermarking for Provenance Tracing and Tamper Localization (AAAI 2026 · better on 2 of 2 shared results)
- BiSLW: Bi-Spectral Latent Watermarking for Generative Diffusion Models (ECCV 2026 · better on 2 of 2 shared results)
- MDDM: Practical Message-Driven Generative Image Steganography Based on Diffusion Models (ICML 2025 · better on 1 of 7 shared results)
- SynTag: Enhancing the Geometric Robustness of Inversion-based Generative Image Watermarking (ICCV 2025 · better on 3 of 3 shared results)
- Your Text Encoder Can Be An Object-Level Watermarking Controller (ICCV 2025 · better on 2 of 2 shared results)
- ROAR: Reducing Inversion Error in Generative Image Watermarking (ICCV 2025 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.88)
- Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models (CVPR 2025 · similarity 0.87)
- CoSDA: Enhancing the Robustness of Inversion-based Generative Image Watermarking Framework (AAAI 2025 · similarity 0.87)
- StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models (NeurIPS 2025 · similarity 0.86)
- A Locally Tokenized Generative Model for Robust Time-Series Watermarking (NeurIPS 2026 · similarity 0.86)
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.
