Research map / Image generation / Autoregressive and latent generation
GoodDiffusion: Proactive Copyright Protection for Diffusion Bridge Models via Learnable Sample-specific Signatures
Shixi Qin, Zhiyong Yang, Shilong Bao, Zitai Wang, Qianqian Xu, Qingming Huang
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 sample-specific signatures (LSN) instead of static signature (ablation) | GoodDiffusion on I2SB · CelebA super-resolution | Abuse Rate under surrogate attack: 98.52 -> 0 | Improved |
| add sample-specific signatures instead of static (ablation) | GoodDiffusion on I2SB · CelebA inpainting | Abuse Rate under surrogate attack: 69.16 -> 0 | Improved |
| add sample-specific signatures instead of static (ablation) | GoodDiffusion on I2SB · CelebA deblurring | Abuse Rate under surrogate attack: 99.97 -> 0.04 | Improved |
| add sample-specific signatures instead of static (ablation) | GoodDiffusion on I2SB · CelebA super-resolution | BFID under surrogate attack: 56.97 -> 261.87 | Improved |
| protect model with GoodDiffusion | DDBM-VP · CelebA super-resolution | FID (authorized): 12.68 -> 16.22 | Hurt |
| protect model with GoodDiffusion | DDBM-VE · CelebA super-resolution | FID (authorized): 13.42 -> 28.64 | Hurt |
| protect model with GoodDiffusion | I2SB · CelebA super-resolution | Abuse Rate: 100 -> 0 | Improved |
| protect model with GoodDiffusion | DBIM · CelebA deblurring | PSNR: 45.40 -> 38.49 | Hurt |
| raise injection strength gamma from 0.9 to 0.99 (weaker signature) (ablation) | GoodDiffusion I2SB · CelebA super-resolution | Abuse Rate: 0 -> 15.88 | Hurt |
| lower gamma from 0.9 to 0.1 (stronger signature) (ablation) | GoodDiffusion I2SB · CelebA super-resolution | FID (authorized): 28.25 -> 57.38 | Hurt |
| fine-tune the protected model for 50 steps (ablation) | GoodDiffusion I2SB · CelebA super-resolution | Abuse Rate: 0 -> 100 | Hurt |
| fine-tune protected model vs train from scratch, 50 steps (ablation) | GoodDiffusion I2SB · CelebA super-resolution | FID: 22.44 -> 75.17 | Hurt |
Closest papers by meaning
- Blameless Users in a Clean Room: Defining Copyright Protection for Generative Models (NeurIPS 2025 · similarity 0.86)
- The Stronger the Diffusion Model, the Easier the Backdoor: Data Poisoning to Induce Copyright BreachesWithout Adjusting Finetuning Pipeline (ICML 2024 · similarity 0.86)
- Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models (CVPR 2025 · similarity 0.85)
- Secure Seed-Based Multi-bit Watermarking for Diffusion Models from First Principles (NeurIPS 2026 · similarity 0.85)
- CPR: Retrieval Augmented Generation for Copyright Protection (CVPR 2024 · similarity 0.85)
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.84)
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.
