Research map / Image generation / Text-to-image diffusion
CLaW: Codec-Guided Adaptive Latent Watermarking for Traceable Diffusion Image Generation
Huangchao Song, Lingyun Yu, Cong Li, Peiqi Jiang, Mingqi Fang, Hongtao Xie
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CLaW, a codec-guided adaptive latent watermarking method for traceable text-to-image diffusion generation that aims to avoid costly retraining and detection while balancing effectiveness and fidelity.
Problem
Traceable watermarking of AI-generated images from text-to-image diffusion models is hard to deploy at scale, due to costly model retraining or detection, poor effectiveness-fidelity trade-offs, and limited (truncated in abstract) capacity or robustness.
Approach
Codec-guided adaptive watermarking operating in the diffusion latent space, intended to enable user-level traceability without retraining the generator. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| inject the watermark only in a late denoising window (ablation) | CLaW · Module ablation (SD3.5-medium) | BitAcc (Adv.): 0.5413 -> 0.9595 | Improved |
| inject the watermark only in a late denoising window (ablation) | CLaW · Module ablation (FLUX.1-dev) | BitAcc (Adv.): 0.5727 -> 0.9695 | Improved |
| add decoder-guided adaptive injection (ablation) | CLaW · Module ablation (SD3.5-medium) | BitAcc (Adv.): 0.9283 -> 0.9595 | Improved |
| add decoder-guided adaptive injection (ablation) | CLaW · Module ablation (FLUX.1-dev) | BitAcc (Adv.): 0.9384 -> 0.9695 | Improved |
| add terminal anchor (ablation) | CLaW · Module ablation (SD3.5-medium) | BitAcc (Adv.): 0.9521 -> 0.9595 | Improved |
| add terminal anchor (ablation) | CLaW · Module ablation (FLUX.1-dev) | BitAcc (Adv.): 0.9604 -> 0.9695 | Improved |
| add decoder-guided adaptive injection (ablation) | CLaW · Module ablation (SD3.5-medium) | PSNR: 42.18 -> 40.04 | Hurt |
| use a late-window start/end instead of earlier windows (ablation) | CLaW · Injection timing sweep | BitAcc and PSNR | Improved |
| raise CFG scale from 5.0 to 10.0 (ablation) | CLaW · SD3.5-medium generation settings | BitAcc (Adv.): 0.9676 -> 0.9480 | Hurt |
| vary denoising steps from 15 to 50 (ablation) | CLaW · SD3.5-medium generation settings | BitAcc (Adv.): 0.9497 -> 0.9582 | No clear effect |
| raise output resolution from 512x512 to 1024x1024 (ablation) | CLaW · SD3.5-medium generation settings | BitAcc (Adv.): 0.9588 -> 0.9704 | No clear effect |
| use CLaW instead of Gaussian Shading/GaussMarker under editing and restoration attacks | CLaW · Composite transformations and restoration (SD3.5-medium) | F1: Editing 0.8293 vs 1.0000 (GaussMarker 0.9975) | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- GaussMarker — GaussMarker: Robust Dual-Domain Watermark for Diffusion Models (ICML 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
- WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.89)
- StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models (NeurIPS 2025 · similarity 0.89)
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Image Watermarks are Removable using Controllable Regeneration from Clean Noise (ICLR 2025 · similarity 0.88)
- CoSDA: Enhancing the Robustness of Inversion-based Generative Image Watermarking Framework (AAAI 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.
