Research map / Image generation / Text-to-image diffusion
Diagnosing and Correcting Concept Omission in Multimodal Diffusion Transformers
Kanghyun Baek, Jaihyun Lew, Chaehun Shin, Jungbeom Lee, Sungroh Yoon
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 Omission Signal Intervention (OSI) | FLUX.1-Dev · GenEval multi-object average | accuracy: 0.47 -> 0.61 | Improved |
| add OSI | SD3.5-Medium · GenEval multi-object average | accuracy: 0.55 -> 0.64 | Improved |
| add OSI to attribute tokens as well as object tokens | FLUX.1-Dev · T2I-CompBench attribute binding (shape) | score: 0.4995 -> 0.5819 | Improved |
| intervene on object tokens only instead of both object and attribute tokens (ablation) | OSI on FLUX · T2I-CompBench attribute binding (color) | score: 0.7923 -> 0.7678 | Hurt |
| reverse intervention direction (opposite) (ablation) | OSI on FLUX · GenEval six-object | accuracy: 0.18 -> 0.02 | Hurt |
| use random direction instead of learned omission direction (ablation) | OSI on FLUX · GenEval six-object | accuracy: 0.18 -> 0.30 (learned: 0.40) | Mixed |
| select top-K probed heads instead of bottom-K heads (ablation) | OSI on FLUX · GenEval five-object | accuracy: 0.22 -> 0.40 | Improved |
| select top-K heads instead of random heads (ablation) | OSI on FLUX · GenEval four-object | accuracy: 0.55 -> 0.64 | Improved |
| select top-K heads instead of all 1368 heads (ablation) | OSI on FLUX · GenEval four-object | accuracy: 0.50 -> 0.64 | Mixed |
| apply OSI to the omitted token vs present tokens (ablation) | OSI on FLUX · 100 failure cases | omitted-object accuracy: 0.14 -> 0.70 | Improved |
| extend intervention from 0 to 5 early steps (ablation) | OSI on FLUX · GenEval two-object | accuracy: 0.82 -> 0.88 | Improved |
| extend intervention from 15 to 30 steps (ablation) | OSI on FLUX · GenEval two-object | accuracy: 0.92 -> 0.91 | No clear effect |
Closest papers by meaning
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.88)
- Rare Text Semantics Were Always There in Your Diffusion Transformer (NeurIPS 2025 · similarity 0.87)
- ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features (ICML 2025 · similarity 0.87)
- Memories of Forgotten Concepts (CVPR 2025 · similarity 0.87)
- EraseAnything: Enabling Concept Erasure in Rectified Flow Transformers (ICML 2025 · similarity 0.87)
- Semantic Surgery: Zero-Shot Concept Erasure in Diffusion 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.
