Research map / Image generation / Text-to-image diffusion
Inference-time T2I improvement
Improve fidelity, safety, and diversity of text-to-image models through noise optimization and guidance.
117 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · 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.
36 papers in 2023–24 and 80 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 9.2% to 9.4%.
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 · Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding (NeurIPS 2022 · 2,090 citations)
- 2023 · Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models (CVPR 2023 · 196 citations)
- 2024 · MMA-Diffusion: MultiModal Attack on Diffusion Models (CVPR 2024 · 53 citations)
- 2025 · SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · 15 citations)
- 2026 · Finding DoRI: Discovery of Retained Images in Diffusion Models (ICML 2026 · 0 citations)
Earliest papers here
- Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding (NeurIPS 2022 · 2,090 citations)
- Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models (CVPR 2023 · 196 citations)
- Stable Diffusion is Unstable (NeurIPS 2023 · 11 citations)
What has been tried
70 papers here report 774 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 7%, hurt 24%, 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 input image, DDIM inversion map, and reconstruction rather than RGB alone | unseen text-to-image generators · AUCROC | 0.410 -> 0.854 (DALL·E 2); 0.399 -> 0.642 (DALL·E 3) | Helped | FakeInversion: Learning to Detect Images from Unseen Text-to-Image Models by Inverting Stable Diffusion (CVPR 2024 · 40 citations) |
| set fingerprint dimension to 128 | Attribution accuracy · attribution accuracy | 0.99 -> 0.94 | Hurt | WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models (CVPR 2024 · 37 citations) |
| Add self-attention conflict loss | Apple-and-pear prompt · subject mixing | Helped | InitNO: Boosting Text-to-Image Diffusion Models via Initial Noise Optimization (CVPR 2024 · 30 citations) | |
| Add the quality-improvement training stage after concept alignment | Generated image quality · Image quality | Helped | AltDiffusion: A Multilingual Text-to-Image Diffusion Model (AAAI 2024 · 26 citations) | |
| use σ = 0.5 instead of σ = 1.0 for EReP data settings | DiffusionDB validation · Aesthetic Score | 5.95 -> 6.01; 5.94 -> 6.03 | Helped | Dynamic Prompt Optimizing for Text-to-Image Generation (CVPR 2024 · 22 citations) |
| Use importance sampling (DDPO_IS) instead of the score function estimator (DDPO_SF) | Compressibility, incompressibility, and aesthetic quality · Reward | Helped | Training Diffusion Models with Reinforcement Learning (ICLR 2024 · 21 citations) | |
| Replace the PH2P optimizer with Adam | COCO · CLIP similarity | 0.77 -> 0.65 | Hurt | Prompting Hard or Hardly Prompting: Prompt Inversion for Text-to-Image Diffusion Models (CVPR 2024 · 19 citations) |
| Increase the trigger sequence length | LoRA rank 640 fine-tuning for 5000 steps · Bit accuracy after fine-tuning | Helped | SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · 15 citations) | |
| Predict the input image x instead of using v-prediction | Surface normals, in-domain · L1 | 0.0514 -> 0.0736 | Hurt | Exploiting Diffusion Prior for Generalizable Dense Prediction (CVPR 2024 · 15 citations) |
| remove dissimilarity guidance (Gsim) | LAION5B text-conditional · % > 0.5 SSCD | 0.00 -> 31.62 | Hurt | Towards Memorization-Free Diffusion Models (CVPR 2024 · 13 citations) |
| Add randomly selected Top5 synthetic captions to LensArt training | LensArt, 250K steps · TIFA | 0.810 -> 0.835 | Helped | On the Scalability of Diffusion-based Text-to-Image Generation (CVPR 2024 · 12 citations) |
| Increase training epochs | General concept categories · Concept fidelity | Helped | Navigating Text-To-Image Customization: From LyCORIS Fine-Tuning to Model Evaluation (ICLR 2024 · 10 citations) |
What reviewers will expect
Datasets and baselines named in the 48 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 · 5 of 48 papers
- DrawBench · 4 of 48 papers
- COCO · 4 of 48 papers
- DiffusionDB · 3 of 48 papers
- MS COCO · 3 of 48 papers
- MS-COCO 2014 · 3 of 48 papers
- I2P · 3 of 48 papers
- T2I-CompBench · 3 of 48 papers
Baselines
- SDXL · 6 of 48 papers
- DDPO · 5 of 48 papers
- Diffusion-DPO · 5 of 48 papers
- CADS · 3 of 48 papers
- DPO · 3 of 48 papers
- SD1.5 · 3 of 48 papers
- ESD · 3 of 48 papers
- UCE · 3 of 48 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 116 papers; from OpenAlex and the papers' affiliations).
- Korea Advanced Institute of Science and Technology · 8 papers
- Google · 6 papers
- The Chinese University of Hong Kong · 5 papers
- Huawei · 4 papers
- Hong Kong University of Science and Technology · 4 papers
- Alibaba Group · 4 papers
- Nanjing University · 4 papers
- Zhejiang University · 3 papers
- Massachusetts Institute of Technology · 3 papers
- Tsinghua University · 3 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
- T2I evaluation and controllable adapters (149 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).