Research map / Image generation / Text-to-image diffusion
ConceptSplit: Decoupled Multi-Concept Personalization of Diffusion Models via Token-wise Adaptation and Attention Disentanglement
Habin Lim, Yeongseob Won, Juwon Seo, Gyeong-Moon Park
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 |
|---|---|---|---|
| train only value projection instead of key and value (ablation) | ToVA · 2 objects, no background | DINO image alignment: 0.474 -> 0.566 | Improved |
| train only value projection instead of key only (ablation) | ToVA · 1 object, no background | DINO image alignment: 0.175 -> 0.809 | Improved |
| add LODA stage 1 latent optimization (ablation) | ConceptSplit, 2 objects · 2 objects, no background | GenEval: 0.550 -> 0.825 | Improved |
| add LODA stage 1 latent optimization (ablation) | ConceptSplit, 3 objects · 3 objects, no background | GenEval: 0.012 -> 0.026 | No clear effect |
| add attention fixing guidance (stage 2) (ablation) | ConceptSplit · 2 objects, no background | GenEval: 0.825 -> 0.902 | Improved |
| add attention fixing guidance (stage 2) (ablation) | ConceptSplit · 3 objects, no background | GenEval: 0.026 -> 0.344 | Improved |
| ConceptSplit versus best baselines on 2 objects with background | multi-concept personalization · 2 objects, with background | GenEval: 0.288 -> 0.648 | Improved |
| ConceptSplit on single object with no background | single-concept personalization · 1 object, no background | GenEval: 0.954 -> 0.942 | No clear effect |
| merging-free token-wise value adapters versus EDLoRA adapter merging | training cost, three concepts · training time | training time: ~84m -> ~9m | Improved |
| modify key projection of cross-attention (ablation) | personalization methods · attention entropy change, dog token | entropy change: -0.003600 -> -0.000266 | Hurt |
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.
- Cones2 — Cones: Concept Neurons in Diffusion Models for Customized Generation (ICML 2023) · better on 2 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
- Attention Calibration for Disentangled Text-to-Image Personalization (CVPR 2024 · similarity 0.91)
- You Can’t Have It Both Ways: Concept Entanglement Limits Diffusion Model Unlearning (NeurIPS 2026 · similarity 0.89)
- How to Continually Adapt Text-to-Image Diffusion Models for Flexible Customization? (NeurIPS 2024 · similarity 0.89)
- Disentangled Sparse Representations for Concept-Separated Diffusion Unlearning (NeurIPS 2026 · similarity 0.88)
- Concept Weaver: Enabling Multi-Concept Fusion in Text-to-Image Models (CVPR 2024 · similarity 0.88)
- Dis²Booth: Learning Image Distribution with Disentangled Features for Text-to-Image Diffusion Models (AAAI 2025 · similarity 0.88)
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.
