Research map / Image generation / Text-to-image diffusion
Unified Safe In-context Image Generation in Multimodal Diffusion Transformers via Restricting Unsafe Information Flows
Xiang Yang, Feifei Li, Mi Zhang, Geng Hong, Xiaoyu You, Mi Wen, Min Yang
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 |
|---|---|---|---|
| apply UVR localized attention regulation instead of existing erasure baselines | FLUX.1-dev · I2P (sexual) T2I | unsafe images detected: 207 -> 40 | Improved |
| apply UVR to adversarial Unsafe-1K prompts | FLUX.1-dev · Unsafe-1K T2I | unsafe images detected: 449 -> 97 | Improved |
| apply UVR to editing with unsafe reference images | FLUX.1-kontext · I2P I2I editing | unsafe images detected: 175 -> 46 | Improved |
| apply UVR to editing: effect on VQA of unsafe-reference edits | FLUX.1-kontext · I2I editing | VQA (%): 81.79 -> 88.42 | Improved |
| apply UVR to T2I generation | FLUX.1-dev · MS-COCO | CLIP score: 31.31 -> 31.32 | No clear effect |
| add spatial connectivity refinement of unsafe masks (ablation) | UVR T2I · Nude T2I | harm (%): 19.9 -> 9.3 | Improved |
| add spatial connectivity refinement of unsafe masks (ablation) | UVR I2I · Nude I2I editing | harm (%): 66.7 -> 25.1 | Improved |
| add feature noise perturbation in the start-up stage (ablation) | UVR T2I · Nude T2I | harm (%): 39.4 -> 9.3 | Improved |
| add feature noise perturbation in the start-up stage (ablation) | UVR T2I · Pikachu T2I | harm (%): 100 -> 27.9 | Improved |
| apply perturbation in all diffusion steps instead of only the start-up stage (ablation) | UVR T2I · Nude and Weapon T2I | harm and CLIP: Nude harm 9.3 -> 0.8; Weapon CLIP 30.92 -> 30.02 | Mixed |
| add attention attenuation toward text tokens for I2I (ablation) | UVR I2I · Nude I2I editing | harm (%): 6.9 -> 25.1 harm, CLIP(u) 22.53 -> 24.21 | Mixed |
| add enhanced Oimg-Oimg self-attention (ablation) | UVR I2I · Nude I2I editing | harm (%): 34.7 -> 25.1 | Improved |
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.
- EraseAnything — EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025) · better on 2 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
- When Understanding Becomes a Risk: Authenticity and Safety Risks in the Emerging Image Generation Paradigm (CVPR 2026 · similarity 0.88)
- Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization (CVPR 2025 · similarity 0.87)
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.86)
- Towards Safe Concept Transfer of Multi-Modal Diffusion via Causal Representation Editing (NeurIPS 2024 · similarity 0.86)
- When Think-with-Image Meets Safety: What Determines Multimodal Jailbreak Robustness? (NeurIPS 2026 · similarity 0.85)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.85)
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.
