Research map / Image generation / Text-to-image diffusion
The Map Is Not the Territory: Embedding-Coverage Blacklists for Safe Diffusion Steering
Juyang Bai, Tong Zhou, Shaolei Ren, Xiaolin Xu
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 MAPS blacklist to SLD-Medium | SLD-Medium · I2P (nudity) | ASR (%): 2.53 -> 1.21 | Improved |
| add MAPS blacklist to SLD-Medium | SLD-Medium · Ring-A-Bell (nudity) | ASR (%): 87.71 -> 65.7 | Improved |
| add MAPS blacklist to TraSCE | TraSCE · Ring-A-Bell (nudity) | ASR (%): 8.6 -> 5.3 | Improved |
| add MAPS blacklist to SAFREE | SAFREE · P4D (nudity) | ASR (%): 38.4 -> 26.12 | Improved |
| add MAPS blacklist to SLD-Weak | SLD-Weak · COCO-30k | FID: 32.80 -> 30.12 | Improved |
| add MAPS blacklist to SLD-Max | SLD-Max · Ring-A-Bell (violence) | ASR (%): 51.20 -> 3.20 | Improved |
| use high-richness long phrases instead of low-richness single words (ablation) | blacklist for SLD · Ring-A-Bell (nudity) | avg ASR (%): 36.15 -> 41.84 | Hurt |
| increase blacklist size from 2 to 6 keywords (ablation) | SLD blacklist · Ring-A-Bell (nudity) | ASR (%): 78.5 -> 76.5 | Improved |
| increase blacklist size from 6 to 8 keywords (ablation) | SLD blacklist · Ring-A-Bell (nudity) | ASR (%): 76.5 -> 80.0 | Hurt |
| restrict original blacklist to nudity-only concepts (ablation) | SLD blacklist · Ring-A-Bell (nudity) | avg ASR (%): 78.59 -> 53.24 | Improved |
| restrict original blacklist to violence-only concepts (ablation) | SLD blacklist · Ring-A-Bell (violence) | avg ASR (%): 75.5 -> 57.9 | Improved |
| replace manual violence blacklist with fine-grained LLM-generated keywords (ablation) | SLD blacklist · Ring-A-Bell (violence) | avg ASR (%): 57.9 -> 46.8 | 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.
- SAFREE — SAFREE: Training-Free and Adaptive Guard for Safe Text-to-Image And Video Generation (ICLR 2025) · better on 5 of 5 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
- Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization (CVPR 2025 · similarity 0.89)
- Training-Free Safe Text Embedding Guidance for Text-to-Image Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Growth Inhibitors for Suppressing Inappropriate Image Concepts in Diffusion Models (ICLR 2025 · similarity 0.87)
- Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails (NeurIPS 2026 · similarity 0.86)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion Models (NeurIPS 2025 · similarity 0.86)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 2024 · similarity 0.86)
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.
