Research map / Image generation / Autoregressive and latent generation
Rel-Zero: Harnessing Patch-Pair Invariance for Robust Zero-Watermarking Against AI Editing
Pengzhen Chen, Yanwei Liu, Xiaoyan Gu, Xiaojun Chen, Wu Liu, Weiping Wang
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 |
|---|---|---|---|
| use ViT-B/16 features instead of ResNet-18 (ablation) | Rel-Zero · ControlNet-Inpainting | TPR@0.1%FPR: 84.13 -> 97.43 | Improved |
| use ViT-B/16 features instead of ResNet-50 (ablation) | Rel-Zero · ControlNet-Inpainting | TPR@0.1%FPR: 85.21 -> 97.43 | Improved |
| add Transformer layers to the MLP pair predictor (ablation) | Rel-Zero · ControlNet-Inpainting | TPR@0.1%FPR: 97.43 -> 92.11 | Hurt |
| add GAT layers to the MLP pair predictor (ablation) | Rel-Zero · ControlNet-Inpainting | TPR@0.1%FPR: 97.43 -> 94.45 | Hurt |
| increase number of top-K pairs up to 50 (ablation) | Rel-Zero · all edit types | robustness | Improved |
| increase patch size from 16x16 to 32x32 (ablation) | Rel-Zero · all edit types | robustness | Hurt |
| use Rel-Zero instead of FGPCET under UltraEdit | zero-watermarking · UltraEdit | TPR@0.1%FPR: 7.25 -> 96.55 | Improved |
| use Rel-Zero instead of ConZWNet under ControlNet-Inpainting | zero-watermarking · ControlNet-Inpainting | TPR@0.1%FPR: 2.41 -> 97.43 | Improved |
| use Rel-Zero instead of VINE under deterministic regeneration | watermarking · Deterministic Regeneration | TPR@0.1%FPR: 99.98 -> 85.13 | Hurt |
| use Rel-Zero instead of VINE under Cropout | watermarking · Cropout | TPR@0.1%FPR: 54.87 -> 98.45 | Improved |
| use Rel-Zero instead of embedding watermarks for fidelity | Rel-Zero · fidelity | SSIM/LPIPS: 1.000 SSIM, 0.000 LPIPS | Improved |
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.
- Robust-Wide — Robust-Wide: Robust Watermarking against Instruction-driven Image Editing (ECCV 2024) · better on 3 of 5 shared results
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
- WMCopier: Forging Invisible Watermarks on Arbitrary Images (NeurIPS 2025 · similarity 0.86)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.86)
- Disentangled Style Domain for Implicit $z$-Watermark Towards Copyright Protection (NeurIPS 2024 · similarity 0.86)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.86)
- OmniGuard: Hybrid Manipulation Localization via Augmented Versatile Deep Image Watermarking (CVPR 2025 · similarity 0.85)
- Invisible Image Watermarks Are Provably Removable Using Generative AI (NeurIPS 2024 · 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.
