Research map / Image generation / Text-to-image diffusion
Multi-concept composition
Customize and compose multiple concepts in text-to-image diffusion models.
56 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · an approach in Text-to-image diffusion, keyword cluster “concept · erasure · unlearning”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
23 papers in 2023–24 and 33 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 5.9% to 3.9%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2023 · Multi-Concept Customization of Text-to-Image Diffusion (CVPR 2023 · 627 citations)
- 2024 · Orthogonal Adaptation for Modular Customization of Diffusion Models (CVPR 2024 · 28 citations)
- 2025 · MultiBooth: Towards Generating All Your Concepts in an Image from Text (AAAI 2025 · 11 citations)
- 2026 · When Do Diffusion Models Learn to Generate Multiple Objects? (ICML 2026 · 0 citations)
Earliest papers here
- Multi-Concept Customization of Text-to-Image Diffusion (CVPR 2023 · 627 citations)
- Mix-of-Show: Decentralized Low-Rank Adaptation for Multi-Concept Customization of Diffusion Models (NeurIPS 2023 · 30 citations)
- Inserting Anybody in Diffusion Models via Celeb Basis (NeurIPS 2023 · 12 citations)
What has been tried
38 papers here report 397 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 7%, hurt 23%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Constrain concept weight bases to be orthogonal | Two merged concepts · Identity preservation | Helped | Orthogonal Adaptation for Modular Customization of Diffusion Models (CVPR 2024 · 28 citations) | |
| Add semantic matching | ViCo dataset · IDINO | 0.638 -> 0.683 | Helped | DreamMatcher: Appearance Matching Self-Attention for Semantically-Consistent Text-to-Image Personalization (CVPR 2024 · 24 citations) |
| Remove the separate-and-strengthen loss | Independent Concept1 and Concept2 · image alignment | Hurt | Attention Calibration for Disentangled Text-to-Image Personalization (CVPR 2024 · 23 citations) | |
| Add weighted masks to MRSA | multi-concept composition · concept identity preservation | Helped | FreeCustom: Tuning-Free Customized Image Generation for Multi-Concept Composition (CVPR 2024 · 23 citations) | |
| add concept activation | T2I-CompBench Color · score | 0.5879 -> 0.7455 | Helped | CoMat: Aligning Text-to-Image Diffusion Model with Image-to-Text Concept Matching (NeurIPS 2024 · 16 citations) |
| Add attribute mappers for pose and background embeddings | DreamBench · CLIP-I (FG) | 0.7563 -> 0.7695 | Helped | Decoupled Textual Embeddings for Customized Image Generation (AAAI 2024 · 14 citations) |
| add the Regional Customization Module | Multi-concept · CLIP-I | 0.691 -> 0.713 | Helped | MultiBooth: Towards Generating All Your Concepts in an Image from Text (AAAI 2025 · 11 citations) |
| Add semantic inversion only | Composition with pretrained concepts · CoI likelihood | 0.032 -> 0.121 | Helped | Compositional Inversion for Stable Diffusion Models (AAAI 2024 · 11 citations) |
| add composed prompt learning | Portrait customization · CSIM | 0.429 -> 0.566 | Helped | MagiCapture: High-Resolution Multi-Concept Portrait Customization (AAAI 2024 · 11 citations) |
| Add transformed objects and extracted form patterns to LLM prompt generation | Small-scale dataset · HE | 3.34 -> 3.52 | Helped | Text-to-Image Generation for Abstract Concepts (AAAI 2024 · 11 citations) |
| add structural feature injection | Multi-concept generation · Text sim | 0.3489 -> 0.3804 | Helped | Concept Weaver: Enabling Multi-Concept Fusion in Text-to-Image Models (CVPR 2024 · 11 citations) |
| Change the classifier used as the concept-alignment oracle | ObjectsImageNet · Pearson correlation with human scores | -0.9892; -0.9888; -0.9816; -0.9763 | No clear effect | ConceptBed: Evaluating Concept Learning Abilities of Text-to-Image Diffusion Models (AAAI 2024 · 7 citations) |
What reviewers will expect
Datasets and baselines named in the 20 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CustomConcept101 · 4 of 20 papers
- DreamBooth dataset · 2 of 20 papers
- I2P · 2 of 20 papers
- Ring-A-Bell · 2 of 20 papers
- MMA · 2 of 20 papers
- ComposLoRA · 2 of 20 papers
- CelebA · 2 of 20 papers
Baselines
- Custom Diffusion · 6 of 20 papers
- Cones 2 · 4 of 20 papers
- Mix-of-Show · 4 of 20 papers
- Textual Inversion · 3 of 20 papers
- LoRA Merge · 3 of 20 papers
- Naive · 2 of 20 papers
- LoRA Switch · 2 of 20 papers
- LoRA Composite · 2 of 20 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 56 papers; from OpenAlex and the papers' affiliations).
- Google · 4 papers
- Adobe · 3 papers
- Zhejiang University · 3 papers
- Bar-Ilan University · 2 papers
- NVIDIA · 2 papers
- OriginAI · 2 papers
- Korea Advanced Institute of Science and Technology · 2 papers
- Hong Kong University of Science and Technology · 2 papers
- Tsinghua University · 2 papers
- Virginia Tech · 2 papers
Use it
Download BibTeX of 24 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Concept erasure in diffusion models (76 papers)
Elsewhere in the topic
- Compositional text-to-image guidance (194 papers)
- Layout-to-image control (164 papers)
- T2I evaluation and controllable adapters (149 papers)
- Text-driven synthesis and semantic guidance (137 papers)
- Personalized and region-controlled generation (136 papers)
- Inference-time T2I improvement (117 papers)
- Multi-subject personalization (82 papers)
- Preference optimization for diffusion (82 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).