Research map / Image generation / Text-to-image diffusion
Automated Red Teaming for Text-to-Image Models through Feedback-Guided Prompt Iteration with Vision-Language Models
Wei Xu, Kangjie Chen, Jiawei Qiu, Yuyang Zhang, Run Wang, Jin Mao, Tianwei Zhang, Lina Wang
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 FGPI single-turn instead of ART | FGPI · Lexica seed set (560 prompts), SD 2.1 | ASR (%): 56.43% -> 77.86% | Improved |
| use FGPI instead of FLIRT for prompt safety | FGPI · Lexica seed set (560 prompts), SD 2.1 | safe prompt rate (%): 54.64% -> 96.96% | Improved |
| use FGPI multi-turn instead of Groot | FGPI · NSFW adversarial prompt set (231 prompts), SD 2.1 | ASR (%): 55.41% -> 78.35% | Improved |
| use FGPI multi-turn instead of SneakyPrompt | FGPI · NSFW adversarial prompt set (231 prompts), SD 2.1 | ASR (%): 62.33% -> 78.35% | Improved |
| use multi-turn instead of single-turn feedback-guide-rewrite (ablation) | FGPI · Lexica seed set (560 prompts), SD 2.1 | ASR (%): 77.86% -> 75.71% | No clear effect |
| use multi-turn instead of single-turn feedback-guide-rewrite (ablation) | FGPI · NSFW adversarial prompt set (231 prompts), SD 2.1 | ASR (%): 72.29% -> 78.35% | Improved |
| add iterative optimization rounds (round 0 to 5) (ablation) | FGPI · Seed-free generation, SD 2.1 | ASR (%): 38.21% -> 88.57% | Improved |
| add iterative optimization rounds (round 0 to 3) (ablation) | FGPI · Seed-free generation, DALL-E 3 | ASR (%): 56.43% -> 85.71% | Improved |
| add iterative optimization rounds (round 0 to 3) (ablation) | FGPI · Seed-free generation, DALL-E 3 | request rejection rate (%): 12.14% -> 6.43% | Improved |
| add iterative optimization rounds (round 0 to 5) (ablation) | FGPI · Seed-free generation, SD 2.1 | safe prompt rate (%): 89.64% -> 97.86% | Improved |
| remove seed prompts (seed-free generation) (ablation) | FGPI · Seed-free generation, SD 2.1 | ASR (%): 77.86% -> 88.57% | Improved |
| transfer prompts from SD 1.5 to SD 2.1 | FGPI · Transfer attack, source SD 1.5 to target SD 2.1 | ASR (%): 100.0 -> 72.79 | Hurt |
Closest papers by meaning
- ART: Automatic Red-teaming for Text-to-Image Models to Protect Benign Users (NeurIPS 2024 · similarity 0.92)
- Red-Teaming Text-to-Image Systems by Rule-based Preference Modeling (NeurIPS 2025 · similarity 0.90)
- RISE: Red-teaming via Iterative Strategy Evolution for Modern Text-to-Image Models (NeurIPS 2026 · similarity 0.90)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.89)
- TRUST-VLM: Thorough Red-Teaming for Uncovering Safety Threats in Vision-Language Models (ICML 2025 · similarity 0.89)
- Chain of Attack: On the Robustness of Vision-Language Models Against Transfer-Based Adversarial Attacks (CVPR 2025 · similarity 0.88)
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.
