Research map / Image generation / Autoregressive and latent generation
SERUM: Simple, Efficient, Robust, and Unifying Marking for Diffusion-based Image Generation
Jan Kociszewski, Hubert Jastrzębski, Tymoteusz Stępkowski, Filip Manijak, Krzysztof Rojek, Franziska Boenisch, Adam Dziedzic
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 dynamic augmentation sampler (m=4) on top of precomputed augmentations (ablation) | SERUM detector · Perturbation robustness avg, SD 2.1 | TPR @ 1% FPR: 97.16 -> 99.75 | Improved |
| add augmentations during detector training (precomputed) (ablation) | SERUM detector · Perturbation robustness avg, SD 2.1 | TPR @ 1% FPR: 78.68 -> 97.16 | Improved |
| increase watermark strength alpha from 0.1 to 0.5 (ablation) | SERUM · Perturbation robustness avg, SD 2.1 | TPR @ 1% FPR: 78.49 -> 99.75 | Improved |
| increase watermark strength alpha from 0.5 to 0.7 (ablation) | SERUM · COCO FID, SD 2.1 | FID: 19.24 -> 20.82 | Hurt |
| set sampler temperature to 1 versus 0 or infinity (ablation) | SERUM augmentation sampler · Perturbation robustness avg, SD 2.1 | TPR @ 1% FPR: 99.48 (tau=0), 99.34 (tau=inf) -> 99.75 (tau=1) | Improved |
| increase sampler batch size m from 4 to 32 (ablation) | SERUM · Perturbation robustness avg, SD 2.1 | TPR @ 1% FPR: 99.75 -> 99.82; training 9.5h -> 60h | No clear effect |
| train detector with perturbation terms removed (ablation) | SERUM · Regen-30 attack | TPR @ 1% FPR: 88.50 -> 52.28 | Hurt |
| omit rotation from training augmentations (ablation) | SERUM detector · Rotate | TPR @ 1% FPR: 99.34 -> 18.46 | Hurt |
| use SERUM watermark instead of GaussMarker/RingID | Stable Diffusion · COCO 2014 FID avg | FID: 20.16 (GaussMarker), 19.70 (RingID) -> 18.96 | Improved |
| use SERUM instead of RingID/GaussMarker on strict-FPR evaluation | Stable Diffusion · Perturbation robustness avg, SD 2.1 | TPR @ 0% FPR: 23.59 (GaussMarker) -> 96.34 | Improved |
| apply SERUM to DiT-based SD 3.0 instead of UNet SD 2.1 | SERUM · Perturbation robustness avg | TPR @ 1% FPR: 99.75 -> 95.24 | Hurt |
| fine-tune on watermarked images with LoRA instead of full fine-tuning | SD 1.4 · Radioactivity | TPR @ 1% FPR: 77.12 -> 52.30 | 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.
- GaussMarker — GaussMarker: Robust Dual-Domain Watermark for Diffusion Models (ICML 2025) · better on 4 of 4 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
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.88)
- Attack-Resilient Image Watermarking Using Stable Diffusion (NeurIPS 2024 · similarity 0.88)
- Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models (CVPR 2025 · similarity 0.88)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.88)
- AquaLoRA: Toward White-box Protection for Customized Stable Diffusion Models via Watermark LoRA (ICML 2024 · 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.
