Research map / Image generation / Text-to-image diffusion
Concept-TRAK: Understanding how diffusion models learn concepts through concept attribution
Yong-Hyun Park, Chieh-Hsin Lai, Satoshi Hayakawa, Yuhta Takida, Naoki Murata, Wei-Hsiang Liao, Woosung Choi, Kin Wai Cheuk, Junghyun Koo, Yuki Mitsufuji
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 |
|---|---|---|---|
| use concept-level local attribution instead of image-level attribution on unseen concept combinations | Concept-TRAK · Synthetic dataset, out-of-distribution | Precision@10: baselines <= 0.50 -> 0.85 | Improved |
| use concept-level attribution on CelebA-HQ compositional OOD case | Concept-TRAK vs DAS · CelebA-HQ, out-of-distribution | Precision@10: 0.67 -> 0.97 | Improved |
| concept-level attribution in the in-distribution CelebA-HQ case | Concept-TRAK vs DAS · CelebA-HQ, in-distribution | Precision@10: 0.96 -> 0.92 | No clear effect |
| add concept-aware utility gradients, DPS-based training gradients, DDIM inversion, and gradient normalization progressively to TRAK with DSM loss (ablation) | TRAK · AbC benchmark (48 samples) | Recall@10: full method 0.955 | Improved |
| set-level (global) attribution by averaging utility gradients, compared with baselines on Toy color concept | Concept-TRAK · Set-level attribution, Toy dataset (average) | Precision@10: D-TRAK 1.00 vs Concept-TRAK 0.95 | Hurt |
| average utility gradients over 256 generated images on CelebA-HQ set-level attribution | baselines · Set-level attribution, CelebA-HQ | Precision@10: 1.00 for all methods | No clear effect |
| concept-level local attribution for rare similar concepts (jaguar vs cat) | Concept-TRAK · Similar big cats, local attribution | Precision@10: cat 0.925, jaguar 0.188 | Mixed |
| concept-level attribution on compositional prompts | Concept-TRAK · Compositional common objects and styles, local object | Precision@10: 0.025 -> 0.934 | Improved |
Closest papers by meaning
- Unveiling Concept Attribution in Diffusion Models (NeurIPS 2025 · similarity 0.90)
- Diffusion Attribution Score: Evaluating Training Data Influence in Diffusion Models (ICLR 2025 · similarity 0.89)
- The Hidden Language of Diffusion Models (ICLR 2024 · similarity 0.88)
- Concept Reachability in Diffusion Models: Beyond Dataset Constraints (ICML 2025 · similarity 0.88)
- Unlearning Concepts in Diffusion Model via Concept Domain Correction and Concept Preserving Gradient (AAAI 2025 · similarity 0.87)
- UnlearnCanvas: Stylized Image Dataset for Enhanced Machine Unlearning Evaluation in Diffusion Models (NeurIPS 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.
