Research map / Image generation / Text-to-image diffusion
Mechanistic Dissection of Cross-Attention Subspaces in Text-to-Image Diffusion Models
Jun-Hyun Bae, Wonyong Jo, Jaehyup Lee, Heechul Jung
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 |
|---|---|---|---|
| nullify top-20% concept-related spectral components of OV matrices | Stable Diffusion v1.4 · I2P (nudity) | ASR (%): 25.03 -> 4.24 | Improved |
| nullify top-20% concept-related spectral components | Stable Diffusion v1.4 · Ring-A-Bell K16 (nudity) | ASR (%): 97.89 -> 41.05 | Improved |
| nullify top-20% concept-related spectral components | Stable Diffusion v1.4 · P4D (nudity) | ASR (%): 69.76 -> 18.54 | Improved |
| compare SN against RECE | SN · Ring-A-Bell K16 (nudity) | ASR (%): 10.53 (RECE) vs 41.05 (SN) | Hurt |
| compare SN against SLD-Strong | SN · Ring-A-Bell K16 (nudity) | ASR (%): 18.95 (SLD-Strong) vs 41.05 (SN) | Hurt |
| compare SN against UCE | SN · MMA-Diffusion (nudity) | ASR (%): 41.00 (UCE) vs 17.60 (SN) | Improved |
| compare SN against SAFREE | SN · Ring-A-Bell K16 (nudity) | ASR (%): 65.26 (SAFREE) vs 41.05 (SN) | Improved |
| scale only identified spectral components (10%) instead of top-10% heads (ablation) | Stable Diffusion v2.1 · concept modulation (qualitative) | visual concept control | Improved |
| remove more than 30% of spectral components (ablation) | Spectral Nullification · image quality | image quality | Hurt |
| remove top-15-20% of concept spectral vectors (ablation) | Spectral Nullification · concept removal (qualitative) | concept presence | Improved |
| remove top-10% concept spectral components (ablation) | Stable Diffusion v2.1 · t-SNE of head outputs | base vs concept cluster separation | 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
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.90)
- Cross-Attention Head Position Patterns Can Align with Human Visual Concepts in Text-to-Image Generative Models (ICLR 2025 · similarity 0.89)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.88)
- ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features (ICML 2025 · similarity 0.88)
- CONFORM: Contrast is All You Need for High-Fidelity Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.87)
- CASL: Concept-Aligned Sparse Latents for Interpreting Diffusion Models (NeurIPS 2026 · 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.
