Research map / Image generation / Autoregressive and latent generation
Robust-Wide: Robust Watermarking against Instruction-driven Image Editing
Runyi Hu, Jie Zhang, Ting Xu, Jiwei Li, Tianwei 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 |
|---|---|---|---|
| add PIDSG module (partial gradient through instruction-driven denoising sampling) during training (ablation) | Robust-Wide · InstructPix2Pix edited images | BER (%) with editing, 64 bits: 50.1558 -> 2.6579 | Improved |
| add L_ex2 loss (extract watermark from unedited watermarked image) (ablation) | Robust-Wide · InstructPix2Pix edited images | BER (%) with editing: 50.1219 -> 2.6579 | Improved |
| increase gradient backward steps k from 1 to 3 (ablation) | Robust-Wide · InstructPix2Pix edited images | BER (%): 4.0520 -> 2.6579 | Improved |
| increase lambda1 (embedding feature-space loss weight) to 0.1 (ablation) | Robust-Wide · InstructPix2Pix edited images | BER (%): 2.6579 -> 50.0364 | Hurt |
| increase lambda2 from 0.1 to 1 (ablation) | Robust-Wide · InstructPix2Pix edited images | BER (%): BER 2.6579 -> 0.9674; PSNR 41.9142 -> 35.2577 | Mixed |
| increase watermark length from 64 to 1024 bits (ablation) | Robust-Wide · InstructPix2Pix edited images | BER (%): 2.6579 -> 6.3036 | Hurt |
| apply Robust-Wide trained on InstructPix2Pix to MagicBrush | Robust-Wide · MagicBrush | average BER (%): 2.6579 -> 9.34 (vs 0.96 on ControlNet-InstructPix2Pix) | Mixed |
| embed watermark with Robust-Wide before editing | image editability · InstructPix2Pix edited images | CLIP-I: 0.8402 -> 0.8430 | No clear effect |
| apply brightness change before editing | Robust-Wide · pixel-level distortions | BER (%): 2.6579 -> 12.2907 | Hurt |
Later papers that compared with this method
4 later accepted papers report Robust-Wide as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Rel-Zero: Harnessing Patch-Pair Invariance for Robust Zero-Watermarking Against AI Editing (CVPR 2026 · better on 3 of 5 shared results)
- GenPTW: Latent Image Watermarking for Provenance Tracing and Tamper Localization (AAAI 2026 · better on 0 of 1 shared result)
- Mask Image Watermarking (NeurIPS 2025 · better on 0 of 3 shared results)
- OmniGuard: Hybrid Manipulation Localization via Augmented Versatile Deep Image Watermarking (CVPR 2025 · better on 0 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
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.91)
- Transferable Black-Box One-Shot Forging of Watermarks via Image Preference Models (NeurIPS 2025 · similarity 0.88)
- An Undetectable Watermark for Generative Image Models (ICLR 2025 · similarity 0.87)
- Hidden in the Noise: Two-Stage Robust Watermarking for Images (ICLR 2025 · similarity 0.86)
- RAW: A Robust and Agile Plug-and-Play Watermark Framework for AI-Generated Images with Provable Guarantees (NeurIPS 2024 · similarity 0.86)
- Leveraging Optimization for Adaptive Attacks on Image Watermarks (ICLR 2024 · 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.
