Research map / Image generation / Text-to-image diffusion
The Illusion of High Utility in Safety Alignment of Text-to-Image Diffusion Models
Adeel Yousaf, Soumik Ghosh, James Beetham, Amrit Singh Bedi, Shah Mubarak
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 DES safety alignment instead of base model, measured on structured TIFA rather than FID/CLIPScore | SD v1.4 · TIFA | TIFA accuracy: 76.3 -> 71.6 | Hurt |
| add safety alignment methods (AdvUnlearn, MACE, SafeCLIP) without geometric regularization | SD v1.4 · TIFA | TIFA accuracy: 76.3 -> 63.1 / 62.6 / 60.1 | Hurt |
| add SAGE geometric regularization (ESP + perturbed LSA) to DES (ablation) | DES · TIFA | TIFA accuracy: 71.6 -> 75.4 | Improved |
| add SAGE regularization to DES | DES · Average ASR | average ASR: 1.0 -> 1.2 | No clear effect |
| add Embedding Spread Preservation loss to DES (ablation) | DES · TIFA | TIFA accuracy: 71.6 -> 74.5 | Improved |
| add perturbed Local Structure Alignment loss alone to DES (ablation) | DES · TIFA / ASR | TIFA accuracy and average ASR: TIFA 71.6 -> 76.0; ASR 1.0 -> 1.7 | Mixed |
| add both ESP and perturbed LSA versus ESP alone (ablation) | DES · CLIPScore | CLIPScore: 26.2 -> 26.4 | Improved |
| use one-sided (directional) ESP penalty instead of symmetric penalty (ablation) | SAGE · Average ASR | average ASR: 2.4 -> 1.1 | Improved |
| add concept perturbation to LSA (ablation) | SAGE · Average ASR | average ASR: 4.3 -> 1.1 | Improved |
| set LSA neighbors K=15 versus other values (ablation) | SAGE · Average ASR | average ASR: 3.8 (K=5) -> 1.1 (K=15) | Improved |
| apply SAGE instead of AlignGuard | SD v2.1 · Average ASR | average ASR: 36.5 -> 12.2 | Improved |
| apply SAGE instead of AlignGuard | SD v2.1 · TIFA | TIFA accuracy: 75.9 -> 75.7 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- AlignGuard — AlignGuard: Scalable Safety Alignment for Text-to-Image Generation (ICCV 2025) · 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
- LLM Alignment--Utility Asymmetry under Semantic-Preserving Transformations (NeurIPS 2026 · similarity 0.88)
- Information Theoretic Text-to-Image Alignment (ICLR 2025 · similarity 0.88)
- A Dense Reward View on Aligning Text-to-Image Diffusion with Preference (ICML 2024 · similarity 0.87)
- Improving Long-Text Alignment for Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.87)
- Confidence-aware Reward Optimization for Fine-tuning Text-to-Image Models (ICLR 2024 · similarity 0.87)
- Prompting4Debugging: Red-Teaming Text-to-Image Diffusion Models by Finding Problematic Prompts (ICML 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.
