Research map / Image generation / Autoregressive and latent generation
Rethinking Tree-Ring Watermarking for Enhanced Multi-Key Identification
Hai Ci, Pei Yang, Yiren Song, Mike Zheng Shou
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 RingID instead of Tree-Ring for verification | Stable Diffusion v2 watermarking · Verification, average AUC | ROC-AUC: 0.975 -> 0.995 | Improved |
| use RingID instead of Tree-Ring for 32-key identification (excluding C&S) | Stable Diffusion v2 watermarking · Identification, 32 keys | identification accuracy: 0.465 -> 0.992 | Improved |
| use RingID instead of Tree-Ring for 2048-key identification (excluding C&S) | Stable Diffusion v2 watermarking · Identification, 2048 keys | identification accuracy: 0.077 -> 0.942 | Improved |
| add discretization of ring values to plus/minus alpha (ablation) | RingID · Identification, 2048 keys | identification accuracy (avg): 0.427 -> 0.819 | Improved |
| add spatial shift to avoid corner cropping under rotation (ablation) | RingID · Identification, 2048 keys, rotation | identification accuracy: 0.000 -> 0.860 | Improved |
| add spatial shift (ablation) | RingID · Identification, 2048 keys, brightness | identification accuracy: 0.940 -> 0.870 | Hurt |
| add lossless imprinting (ablation) | RingID · Identification, 2048 keys | identification accuracy (avg): 0.700 -> 0.819 | Improved |
| draw a rounder ring (ablation) | RingID · Identification, 2048 keys, rotation | identification accuracy: 0.620 -> 0.860 | Improved |
| add Gaussian noise watermark on channel 0 alongside ring watermark (ablation) | RingID · Identification, 2048 keys | identification accuracy (avg): 0.740 -> 0.819 | Improved |
| imprint noise watermark on channels 0 and 1 instead of channel 0 only (ablation) | RingID · Identification, 2048 keys | identification accuracy (avg): 0.819 -> 0.830 | Improved |
| imprint noise watermark on channels 0, 1, 2 (ablation) | RingID · Identification, 2048 keys | identification accuracy (avg): 0.830 -> 0.807 | Hurt |
| use RingID vs Tree-Ring under crop and scale for identification | RingID · Identification, 32 to 2048 keys | identification accuracy: 0.530 -> 0.080 (RingID, 32 to 2048 keys) | No clear effect |
Closest papers by meaning
- Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models (CVPR 2025 · similarity 0.86)
- Secure Seed-Based Multi-bit Watermarking for Diffusion Models from First Principles (NeurIPS 2026 · similarity 0.86)
- Can Simple Averaging Defeat Modern Watermarks? (NeurIPS 2024 · similarity 0.84)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.84)
- Leveraging Optimization for Adaptive Attacks on Image Watermarks (ICLR 2024 · similarity 0.83)
- Watermarks in the Sand: Impossibility of Strong Watermarking for Language Models (ICML 2024 · similarity 0.83)
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.
