Research map / Image generation / Autoregressive and latent generation
ClusterMark: Towards Robust Watermarking for Autoregressive Image Generators with Visual Token Clustering
Denis Lukovnikov, Andreas Müller, Erwin Quiring, Asja Fischer
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 |
|---|---|---|---|
| cluster green/red sets (k=64) instead of per-token sets (ablation) | KGW-style AR watermark (LlamaGen GPT-B) · Gaussian noise std .2 | TPR@FPR=1%: 0.075 -> 0.369 | Improved |
| cluster green/red sets (k=64) instead of per-token sets (ablation) | KGW-style AR watermark (RAR-XL) · Gaussian blur R3 | TPR@FPR=1%: 0.045 -> 0.100 | Improved |
| cluster green/red sets (k=64) (ablation) | KGW-style AR watermark (LlamaGen GPT-L) · Regeneration attack | TPR@FPR=1%: 0.812 -> 0.997 | Improved |
| add cluster classifier to clustered watermark (ablation) | Ours (Clustering, k=64), LlamaGen GPT-L · Salt & Pepper .1 | TPR@FPR=1%: 0.543 -> 1.000 | Improved |
| add token classifier to clustered watermark (ablation) | Ours (Clustering, k=64), RAR-XL · Gaussian blur R3 | TPR@FPR=1%: 0.100 -> 0.837 | Improved |
| add token classifier without clustering (ablation) | Ours (No Clustering), LlamaGen GPT-B · Gaussian blur R3 | TPR@FPR=1%: 0.068 -> 0.366 | Improved |
| add clustering (ablation) | Ours, LlamaGen GPT-L · ImageNet FID | FID: 4.36 -> 4.85 | Hurt |
| add clustering (ablation) | Ours, RAR-XL · ImageNet FID | FID: 3.14 -> 3.96 | Hurt |
| lower number of clusters (below 64) (ablation) | Training-free clustered watermark · TPR and FID (Fig. 3, 4) | TPR@FPR=1%, FID: k=8 to 128 | Mixed |
| higher penalty delta (2 -> 5) (ablation) | Clustered watermark · TPR under perturbations | TPR@FPR=1% | Improved |
| lower green fraction (0.5 -> 0.25) (ablation) | Clustered watermark · TPR / FID | TPR@FPR=1%, FID | Mixed |
| choose best hash prefix (kappa=5 vs kappa=1) (ablation) | Clustered watermark k=8, LlamaGen GPT-B · green token ratio of unwatermarked images | separability | Improved |
Closest papers by meaning
- Watermarking Autoregressive Image Generation (NeurIPS 2025 · similarity 0.91)
- BitMark: Watermarking Bitwise Autoregressive Image Generative Models (NeurIPS 2025 · similarity 0.90)
- A Locally Tokenized Generative Model for Robust Time-Series Watermarking (NeurIPS 2026 · similarity 0.89)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.88)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.88)
- Robust Watermarking Using Generative Priors Against Image Editing: From Benchmarking to Advances (ICLR 2025 · similarity 0.87)
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.
