Research map / Image generation / Text-to-image diffusion
T2I evaluation and controllable adapters
Benchmark text-to-image models with human feedback and add controllable adapters.
149 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Text-to-image diffusion, keyword cluster “t2i · preference · alignment”
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.
26 papers in 2023–24 and 121 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 6.6% to 14.2%.
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
- 2020 · Reinforcing an Image Caption Generator Using Off-Line Human Feedback (AAAI 2020 · 25 citations)
- 2022 · Mutual Information Divergence: A Unified Metric for Multimodal Generative Models (NeurIPS 2022 · 7 citations)
- 2023 · GenEval: An object-focused framework for evaluating text-to-image alignment (NeurIPS 2023 · 80 citations)
- 2024 · T2I-Adapter: Learning Adapters to Dig Out More Controllable Ability for Text-to-Image Diffusion Models (AAAI 2024 · 865 citations)
- 2025 · LLM4GEN: Leveraging Semantic Representation of LLMs for Text-to-Image Generation (AAAI 2025 · 19 citations)
- 2026 · EvalMuse-40K: A Fine-Grained Benchmark with Comprehensive Human Annotations for Text-to-Image Generation Model Alignment Evaluation (AAAI 2026 · 3 citations)
Earliest papers here
- Reinforcing an Image Caption Generator Using Off-Line Human Feedback (AAAI 2020 · 25 citations)
- Mutual Information Divergence: A Unified Metric for Multimodal Generative Models (NeurIPS 2022 · 7 citations)
- GenEval: An object-focused framework for evaluating text-to-image alignment (NeurIPS 2023 · 80 citations)
What has been tried
64 papers here report 704 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 7%, hurt 25%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Inject guidance into the UNet encoder instead of the decoder at 4 scales | COCO validation set, sketch guidance · MSE | 0.1455 -> 0.1280 | Helped | T2I-Adapter: Learning Adapters to Dig Out More Controllable Ability for Text-to-Image Diffusion Models (AAAI 2024 · 865 citations) |
| Use task-augmented prompts instead of separate heads for plausibility prediction | RichHF-18K test set: plausibility · PLCC | 0.666 -> 0.693 | Helped | Rich Human Feedback for Text-to-Image Generation (CVPR 2024 · 63 citations) |
| Fine-tune both the text identifier and U-Net instead of only the identifier | CUB, 5-shot · accuracy | 57.64 -> 59.41 | Helped | Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model (CVPR 2024 · 35 citations) |
| Use Llama-2/7B as the LLM | T2I-CompBench · Color | 37.65 -> 43.21 | Helped | LLM4GEN: Leveraging Semantic Representation of LLMs for Text-to-Image Generation (AAAI 2025 · 19 citations) |
| Use top-1 rather than top-3 retrieved prompts | PIP test set (short sentence) · PMS | 0.6179 -> 0.6057 | Hurt | Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting (CVPR 2024 · 17 citations) |
| Change the LLM used for unsafe-word selection and substitution among GPT3.5, GPT4o, and Tongyiqianwen | discrimination, illegal, pornographic, privacy, and violent prompts · ASR | 0.830, 0.910, 0.890; 0.980, 0.980, 0.980; 0.485, 0.605, 0.795; 0.910, 0.905, 0.970; 0.645, 0.780, 0.855 | Mixed | Perception-Guided Jailbreak Against Text-to-Image Models (AAAI 2025 · 11 citations) |
| discriminatively tune with global matching | COCO-NSS1K (ID) · CLIP | 34.96 -> 35.14 | Helped | Discriminative Probing and Tuning for Text-to-Image Generation (CVPR 2024 · 11 citations) |
| Use all four reward models rather than CLIPScore alone | T2I-CompBench attribute binding, first 50 prompts · Shape | 66.9 to 77.4 | Helped | ReNO: Enhancing One-step Text-to-Image Models through Reward-based Noise Optimization (NeurIPS 2024 · 11 citations) |
| add the Verbalizer to the Sentence-Similarity Checker | SneakyPrompt · AUROC (%) | 97.39 -> 97.86 | Helped | GuardT2I: Defending Text-to-Image Models from Adversarial Prompts (NeurIPS 2024 · 6 citations) |
| Add a Unet-style long connection | T2I framework training · training convergence speed | four times faster | Helped | DreamOmni: Unified Image Generation and Editing (CVPR 2025 · 4 citations) |
| Add DOTS while omitting SWFR | UltraHR-eval4K · FID | 33.99 -> 32.57 | Helped | UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (NeurIPS 2025 · 4 citations) |
| Use diverse or deterministic language-enhanced prompts instead of minimal prompts | PACS, OfficeHome, and DomainNet SDG · accuracy | 76.43 -> 75.65; 64.66 -> 63.14; 31.94 -> 30.94 | Hurt | Not Just Pretty Pictures: Toward Interventional Data Augmentation Using Text-to-Image Generators (ICML 2024 · 3 citations) |
What reviewers will expect
Datasets and baselines named in the 46 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- PartiPrompts · 7 of 46 papers
- MSCOCO · 5 of 46 papers
- ImageNet-1K · 3 of 46 papers
- ADE20K · 3 of 46 papers
- T2I-CompBench · 2 of 46 papers
- GenAI-Bench · 2 of 46 papers
- LAION · 2 of 46 papers
- GenEval · 2 of 46 papers
Baselines
- SDXL · 6 of 46 papers
- FLUX.1[dev] · 4 of 46 papers
- PixArt-α · 3 of 46 papers
- Sana · 3 of 46 papers
- ControlNet · 3 of 46 papers
- CLIP · 3 of 46 papers
- VQAScore · 3 of 46 papers
- GORS · 2 of 46 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 147 papers; from OpenAlex and the papers' affiliations).
- Google · 8 papers
- Adobe · 4 papers
- Zhejiang University · 4 papers
- Nanyang Technological University · 4 papers
- The Chinese University of Hong Kong · 4 papers
- Shanghai Jiao Tong University · 4 papers
- Hong Kong University of Science and Technology · 3 papers
- Stanford University · 3 papers
- University of Washington · 3 papers
- Tsinghua University · 3 papers
Use it
Download BibTeX of 27 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
- Inference-time T2I improvement (117 papers)
- Preference optimization for diffusion (82 papers)
- Fairness and bias in generative models (52 papers)
Elsewhere in the topic
- Compositional text-to-image guidance (194 papers)
- Layout-to-image control (164 papers)
- Text-driven synthesis and semantic guidance (137 papers)
- Personalized and region-controlled generation (136 papers)
- Multi-subject personalization (82 papers)
- Concept erasure in diffusion models (76 papers)
- Text generation and structured prediction (65 papers)
- Multi-concept composition (56 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).