Research map / Image generation / Text-to-image diffusion
Compositional text-to-image guidance
Improve object composition in text-to-image diffusion through attention guidance and token-level supervision.
194 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in Text-to-image diffusion, keyword cluster “synthesis · layout · object”
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.
87 papers in 2023–24 and 105 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 22.3% to 12.3%.
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
- 2022 · GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models (ICML 2022 · 989 citations)
- 2023 · DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation (CVPR 2023 · 2,183 citations)
- 2024 · SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis (ICLR 2024 · 322 citations)
- 2025 · Diffusion-4K: Ultra-High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2025 · 20 citations)
- 2026 · EasyText: Controllable Diffusion Transformer for Multilingual Text Rendering (AAAI 2026 · 4 citations)
Earliest papers here
- GLIDE: Towards Photorealistic Image Generation and Editing with Text-Guided Diffusion Models (ICML 2022 · 989 citations)
- Vector Quantized Diffusion Model for Text-to-Image Synthesis (CVPR 2022 · 675 citations)
- DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation (CVPR 2023 · 2,183 citations)
What has been tried
127 papers here report 1,376 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 6%, hurt 20%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add the refinement stage to SDXL | User study · win rate | 36.93% -> 48.44% | Helped | SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis (ICLR 2024 · 322 citations) |
| Add instance-level captions during training | COCO val instance attribute binding · Color accuracy | 38.2(17.2) -> 55.4 | Helped | InstanceDiffusion: Instance-level Control for Image Generation (CVPR 2024 · 82 citations) |
| Use SAR without CAR | HRS · Spatial accuracy | 30.74 -> 36.43 | Helped | Grounded Text-to-Image Synthesis with Attention Refocusing (CVPR 2024 · 74 citations) |
| Add disentangled conditioning mechanism (DCM) | Evaluation set · text alignment | 0.148 -> 0.224 | Helped | DEADiff: An Efficient Stylization Diffusion Model with Disentangled Representations (CVPR 2024 · 72 citations) |
| Replace cross-attention maps during editing | Image editing examples in Figure 4 · Editing success | Hurt | Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing (CVPR 2024 · 72 citations) | |
| use the UFOGen objective instead of the SIDDM objective for one-step sampling | MSCOCO-2017 5k · FID-5k | 28.0 -> 22.5 | Helped | UFOGen: You Forward Once Large Scale Text-to-Image Generation via Diffusion GANs (CVPR 2024 · 59 citations) |
| Remove structure guidance | Controllable T2I diffusion · structure alignment | Hurt | FreeControl: Training-Free Spatial Control of Any Text-to-Image Diffusion Model with Any Condition (CVPR 2024 · 42 citations) | |
| Add joint initialization of panorama and perspective latent maps | Matterport3D panoramas · FID | 59.43 -> 46.47 | Helped | Taming Stable Diffusion for Text to 360 Panorama Image Generation (CVPR 2024 · 37 citations) |
| Add BoxNet and non-unique cross-attention mask control | COCO · Min. Object Score | 0.3973 ± 0.0021 -> 0.4010 ± 0.0028 | No clear effect | Compositional Text-to-Image Synthesis with Attention Map Control of Diffusion Models (AAAI 2024 · 35 citations) |
| replace residual cross-attention with parallel cross-attention | Personalized generation evaluation · CLIP Score | .267 -> .243 | Hurt | When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation (CVPR 2024 · 35 citations) |
| add localized attention constraint | Ablation evaluation · CLIP | 0.8065 -> 0.8108 | Helped | Brush Your Text: Synthesize Any Scene Text on Images via Diffusion Model (AAAI 2024 · 25 citations) |
| Use GPT-4o captions instead of LAION-5B captions | Aesthetic-Eval@4096 · CLIPScore | 29.37 -> 33.12 | Helped | Diffusion-4K: Ultra-High-Resolution Image Synthesis with Latent Diffusion Models (CVPR 2025 · 20 citations) |
What reviewers will expect
Datasets and baselines named in the 57 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CC12M · 4 of 57 papers
- NYUv2 · 4 of 57 papers
- SA-1B · 3 of 57 papers
- T2I-CompBench · 3 of 57 papers
- MSCOCO · 3 of 57 papers
- COCO · 3 of 57 papers
- ImageNet · 3 of 57 papers
- Hypersim · 3 of 57 papers
Baselines
- Marigold · 5 of 57 papers
- Stable Diffusion · 4 of 57 papers
- ControlNet · 4 of 57 papers
- PixArt-Σ · 4 of 57 papers
- CONFORM · 4 of 57 papers
- GeoWizard · 4 of 57 papers
- Structured Diffusion · 3 of 57 papers
- T2I-Adapter · 3 of 57 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 192 papers; from OpenAlex and the papers' affiliations).
- Google · 12 papers
- Korea Advanced Institute of Science and Technology · 10 papers
- Peking University · 8 papers
- Zhejiang University · 8 papers
- Alibaba Group · 7 papers
- Tsinghua University · 7 papers
- Meta · 6 papers
- Microsoft · 5 papers
- The University of Hong Kong · 5 papers
- Google DeepMind · 5 papers
Use it
Download BibTeX of 34 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
- Text-driven synthesis and semantic guidance (137 papers)
- Personalized and region-controlled generation (136 papers)
- Layout-to-image control (164 papers)
- Multi-subject personalization (82 papers)
- Text generation and structured prediction (65 papers)
Elsewhere in the topic
- T2I evaluation and controllable adapters (149 papers)
- Inference-time T2I improvement (117 papers)
- Preference optimization for diffusion (82 papers)
- Concept erasure in diffusion models (76 papers)
- Multi-concept composition (56 papers)
- Fairness and bias in generative models (52 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).