Research map / Image generation / Text-to-image diffusion
Hidden Dangers of Compositional Generation: Diagnosing Semantic Safety Failures in Text-to-Image Models
Haoming Yang, Ke Ma, Ligong Zhang, Xiaojun Jia, Yingfei Sun, Qianqian Xu, Qingming Huang
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 fine-grained decomposition and reassembly on top of visual metaphors (ablation) | CoRA · DALL-E 3 | ASR: 0.081 -> 0.644 | Improved |
| add decomposition and reassembly on top of visual metaphors (ablation) | CoRA · Cogview4 | ASR: 0.444 -> 0.733 | Improved |
| add decomposition and reassembly on top of visual metaphors (ablation) | CoRA · Cogview4 | Prompt perplexity: 97.193 -> 37.278 | Improved |
| add decomposition and reassembly on top of visual metaphors (ablation) | CoRA · DALL-E 3 | Inception Score: 1.000 -> 3.813 | Improved |
| use stronger auxiliary model Qwen3-235B instead of Qwen3-8B (ablation) | CoRA · five T2I models | ASR: 0.73 -> 0.74 (Cogview4) | No clear effect |
| use weaker auxiliary model Qwen2-7B instead of Qwen3-8B (ablation) | CoRA · SDXL | ASR: 0.59 -> 0.57 | No clear effect |
| use CoRA instead of DACA | T2I safety attack · DALL-E 3 | ASR: 0.407 -> 0.644 | Improved |
| use CoRA instead of DACA | T2I safety attack · GPT-4o web interface | ASR: 0.433 -> 0.667 | Improved |
| use CoRA instead of MMA-Diffusion on DALL-E 3 semantic consistency | T2I safety attack · DALL-E 3 | Semantic Consistency: 0.261 -> 0.258 | No clear effect |
| use CoRA instead of Ring-a-Bell on prompt perplexity | T2I safety attack · Cogview4 | Prompt perplexity: 12752.81 -> 37.28 | 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.
- MMA-Diffusion — MMA-Diffusion: MultiModal Attack on Diffusion Models (CVPR 2024) · better on 5 of 6 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
- ART: Automatic Red-teaming for Text-to-Image Models to Protect Benign Users (NeurIPS 2024 · similarity 0.89)
- When Understanding Becomes a Risk: Authenticity and Safety Risks in the Emerging Image Generation Paradigm (CVPR 2026 · similarity 0.88)
- T2ISafety: Benchmark for Assessing Fairness, Toxicity, and Privacy in Image Generation (CVPR 2025 · similarity 0.87)
- T2VSafetyBench: Evaluating the Safety of Text-to-Video Generative Models (NeurIPS 2024 · similarity 0.87)
- UPAM: Unified Prompt Attack in Text-to-Image Generation Models Against Both Textual Filters and Visual Checkers (ICML 2024 · similarity 0.87)
- OVERT: A Benchmark for Over-Refusal Evaluation on Text-to-Image Models (NeurIPS 2025 · 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.
