Research map / Image generation / Text-to-image diffusion
SafeRoPE: Risk-specific Head-wise Embedding Rotation for Safe Generation in Rectified Flow Transformers
Xiang Yang, Feifei Li, Mi Zhang, Geng Hong, Xiaoyu You, 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 SafeRoPE risk-specific head-wise rotation | FLUX.1-dev · Nudity, I2P | Unsafe rate (%): 10.3 -> 7.0 | Improved |
| apply SafeRoPE | FLUX.1-dev · Nudity, Unsafe-1K | Unsafe rate (%): 38.8 -> 15.4 | Improved |
| apply SafeRoPE | FLUX.1-dev · Bloody concept | Unsafe rate (%): 68.1 -> 15.5 | Improved |
| apply SafeRoPE | FLUX.1-dev · COCO-1K utility | VQA score: 87.5 -> 89.7 | Improved |
| apply SafeRoPE | FLUX.1-dev · COCO-1K utility | CLIP score: 31.3 -> 31.3 | No clear effect |
| apply SafeRoPE | FLUX.1-dev · COCO-1K utility | FID: 76.8 -> 68.9 | Improved |
| transfer rotation matrices learned on FLUX.1-dev to FLUX.1-schnell | SafeRoPE · Bloody concept | Unsafe rate (%): 34.2 -> 16.4 | Improved |
| transfer dev-trained rotations rather than training on schnell (ablation) | SafeRoPE · Bloody concept (FLUX.1-schnell) | Unsafe rate (%): 13.7 -> 16.4 | Hurt |
| use learned rotations instead of random rotations (ablation) | SafeRoPE · Nudity, Unsafe-1K | Unsafe rate (%): 35.2 -> 15.4 | Improved |
| use independent rotation matrices with scaled image-token initialization instead of shared/unscaled (ablation) | SafeRoPE · Nudity, Unsafe-1K | Unsafe rate (%): 24.2-29.0 -> 15.4 | Improved |
| raise rotation rank from 2 to default (ablation) | SafeRoPE · Nudity, Unsafe-1K | Unsafe rate (%): 34.0 -> 15.4 | Improved |
| raise rotation rank to 10 instead of default (ablation) | SafeRoPE · Nudity, I2P | Unsafe rate (%): 7.0 -> 11.1 | Hurt |
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
- Safe Few-Step Generation via Velocity Editing (NeurIPS 2026 · similarity 0.88)
- SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025 · similarity 0.88)
- Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Towards Resilient Safety-driven Unlearning for Diffusion Models against Downstream Fine-tuning (NeurIPS 2025 · similarity 0.86)
- When Understanding Becomes a Risk: Authenticity and Safety Risks in the Emerging Image Generation Paradigm (CVPR 2026 · similarity 0.86)
- Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection (CVPR 2026 · 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.
