Research map / Image generation / Text-to-image diffusion
AlignGuard: Scalable Safety Alignment for Text-to-Image Generation
Runtao Liu, I Chieh Chen, Jindong Gu, Jipeng Zhang, Renjie Pi, Qifeng Chen, Philip Torr, Ashkan Khakzar, Fabio Pizzati
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 paired safe image as DPO positive instead of black image (ablation) | AlignGuard SD v1.5 · CoProV2 test (ablation) | IP: 0.21 -> 0.07 | Improved |
| use paired safe image as DPO positive instead of warning sign (ablation) | AlignGuard SD v1.5 · CoProV2 test (ablation) | IP: 0.16 -> 0.07 | Improved |
| use paired safe image as DPO positive instead of black image (ablation) | AlignGuard SD v1.5 · COCO | FID: 85.38 -> 70.96 | Improved |
| add consistency loss L_con (ablation) | AlignGuard SD v1.5 · COCO | FID: 74.82 -> 70.96 | Improved |
| add consistency loss L_con (ablation) | AlignGuard SD v1.5 · COCO | CLIPScore: 30.78 -> 32.32 | Improved |
| add consistency loss L_con (ablation) | AlignGuard SD v1.5 · CoProV2 test (ablation) | IP: 0.16 -> 0.07 | Improved |
| use Co-Merge instead of TIES (ablation) | AlignGuard SD v1.5 · COCO | FID: 79.79 -> 70.96 | Improved |
| use Co-Merge instead of Task Vectors (ablation) | AlignGuard SD v1.5 · COCO | CLIPScore: 4.80 -> 32.32 | Improved |
| use Co-Merge instead of Model Soups (ablation) | AlignGuard SD v1.5 · CoProV2 test (ablation) | IP: 0.14 -> 0.07 | Improved |
| merge per-category experts (Co-Merge) rather than single LoRA on all data (ablation) | AlignGuard SD v1.5 · CoProV2 test | avg IP: 0.16 -> 0.07 | Improved |
| scale training data from 10% to 100% (ablation) | AlignGuard SD v1.5 · CoProV2 test (ablation) | IP: 0.19 -> 0.07 | Improved |
| apply AlignGuard alignment | SDXL · COCO | FID: 68.90 -> 75.20 | Hurt |
Later papers that compared with this method
3 later accepted papers report AlignGuard as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- The Path to Reconciling Quality and Safety Alignment in Text-to-Image Generation (ECCV 2026 · better on 6 of 7 shared results)
- Towards Seed-Robust Safety Alignment in Text-to-Image Models (ICML 2026 · better on 5 of 5 shared results)
- The Illusion of High Utility in Safety Alignment of Text-to-Image Diffusion Models (ECCV 2026 · better on 1 of 2 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
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.89)
- SafeSora: Towards Safety Alignment of Text2Video Generation via a Human Preference Dataset (NeurIPS 2024 · similarity 0.88)
- OVERT: A Benchmark for Over-Refusal Evaluation on Text-to-Image Models (NeurIPS 2025 · similarity 0.88)
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.88)
- T2VSafetyBench: Evaluating the Safety of Text-to-Video Generative Models (NeurIPS 2024 · similarity 0.87)
- GuardT2I: Defending Text-to-Image Models from Adversarial Prompts (NeurIPS 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.
