Research map / Image generation / Text-to-image diffusion
Towards Seed-Robust Safety Alignment in Text-to-Image Models
Zhenyu Wu, Yao Huang, Shouwei Ruan, Xingxing Wei
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 parameter adaptation (ablation) | NCD · NSFW-56K | SSR-10: 24.9% -> 19.7% | Improved |
| add regularization term on top of parameter adaptation (ablation) | NCD · NSFW-56K | SSR-10: 19.7% -> 14.8% | Improved |
| add regularization term (ablation) | NCD · MMA-Diffusion | SSR-10: 32.2% -> 20.0% | Improved |
| add parameter adaptation and regularization (ablation) | NCD · COCO-30K | CLIP score: 26.36 to 26.39 | No clear effect |
| apply basic NCA alignment to SD-v1.5 (ablation) | SD-v1.5 · NSFW-56K | SSR-10: 86.7% to 24.9% | Improved |
| apply NCD to SD-v1.5 | SD-v1.5 · I2P-Sexual | SSR-10: 11.1% to 6.2% | Improved |
| apply NCD to SD-v1.5 | SD-v1.5 · MMA-Diffusion | SSR-10: 94.2% to 20.0% | Improved |
| increase number of seeds from 3 to 50 for ESD-u (ablation) | ESD-u · I2P-Sexual | SSR-N: 25% -> 49% | Hurt |
| increase number of seeds from 3 to 50 for RECE (ablation) | RECE · I2P-Sexual | SSR-N: 11% to 31% | Hurt |
| increase seeds to 50 for NCD (ablation) | NCD · MMA-Diffusion | SSR-50: 36.1% | Mixed |
| use NCD instead of RECE under adversarial seed selection | SD-v1.5 · Adaptive seed selection | SSR-50: 48.6% -> 28.2% | Improved |
| use NCD instead of ESD under latent inversion | SD-v1.5 · Latent inversion attack | ASR: 0.282 -> 0.049 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- AlignGuard — AlignGuard: Scalable Safety Alignment for Text-to-Image Generation (ICCV 2025) · better on 5 of 5 shared results
- Safree — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 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
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.89)
- Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning (NeurIPS 2025 · similarity 0.89)
- Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization (CVPR 2025 · similarity 0.88)
- Leveraging Catastrophic Forgetting to Develop Safe Diffusion Models against Malicious Finetuning (NeurIPS 2024 · similarity 0.88)
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.87)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 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.
