Research map / Image generation / Autoregressive and latent generation
ROAR: Reducing Inversion Error in Generative Image Watermarking
Hanyi Wang, Han Fang, Shi-Lin Wang, Ee-Chien Chang
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 regeneration-based optimization (RO) only (ablation) | Tree-Ring · Gaussian noise sigma=0.2 | TPR: 0.101 -> 0.525 | Improved |
| add distortion-adaptive restoration (AR) only (ablation) | Tree-Ring · Gaussian noise sigma=0.2 | TPR: 0.101 -> 0.857 | Improved |
| add both RO and AR (ROAR) (ablation) | Tree-Ring · Gaussian noise sigma=0.2 | TPR: 0.101 -> 1.000 | Improved |
| add AR only on unseen strong blur (ablation) | Tree-Ring · Gaussian blur r=9 | TPR: 0.687 -> 0.731 | No clear effect |
| add RO for unseen blur (ablation) | Tree-Ring · Gaussian blur r=9 | TPR: 0.687 -> 0.956 | Improved |
| plug ROAR into Gaussian Shading | Gaussian Shading · SD v1.4/2.1, 9 distortions | Bit accuracy (distorted): 0.885 -> 0.988 | Improved |
| plug ROAR into Tree-Ring and RingID | Tree-Ring, RingID · SD v1.4/2.1, 9 distortions | TPR (distorted): 0.838 -> 1.000 (Tree-Ring) | Improved |
| ROAR on image quality | Gaussian Shading · SD | FID: 25.20 -> 25.20 | No clear effect |
| mismatch inversion step vs inference step (ablation) | ROAR with GS · Bit accuracy | bit accuracy: 1.000 -> 0.994 (worst case) | No clear effect |
| increase RO optimization iterations beyond 20 (ablation) | ROAR · TPR / bit accuracy | extraction accuracy | No clear effect |
| stronger purification attack strength | GS-ROAR · purification attack | TPR: 1.000 (s=0.3) -> 0.663 (s=0.7) | Hurt |
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.
- LaWa — LaWa: Using Latent Space for In-Generation Image Watermarking (ECCV 2024) · 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
- CoSDA: Enhancing the Robustness of Inversion-based Generative Image Watermarking Framework (AAAI 2025 · similarity 0.91)
- PGID: Progressive Guided Inversion and Denoising for Robust Watermark Detection (NeurIPS 2026 · similarity 0.87)
- Secure Seed-Based Multi-bit Watermarking for Diffusion Models from First Principles (NeurIPS 2026 · similarity 0.86)
- ROBIN: Robust and Invisible Watermarks for Diffusion Models with Adversarial Optimization (NeurIPS 2024 · similarity 0.86)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.85)
- BitMark: Watermarking Bitwise Autoregressive Image Generative Models (NeurIPS 2025 · similarity 0.85)
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.
