Research map / Image generation / Text-to-image diffusion
Preference optimization for diffusion
Align diffusion models with human preferences using direct preference optimization.
82 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2023–2026) · 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.
8 papers in 2023–24 and 74 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 2.0% to 8.7%.
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 · ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation (NeurIPS 2023 · 108 citations)
- 2024 · Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · 23 citations)
- 2025 · VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · 18 citations)
- 2026 · VisionReward: Fine-Grained Multi-Dimensional Human Preference Learning for Image and Video Generation (AAAI 2026 · 6 citations)
Earliest papers here
- ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation (NeurIPS 2023 · 108 citations)
- Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation (NeurIPS 2023 · 99 citations)
- Rewarded soups: towards Pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards (NeurIPS 2023 · 8 citations)
What has been tried
32 papers here report 344 measured changes (a component added, removed or swapped, compared with a baseline): helped 67%, no clear effect 3%, hurt 24%, mixed 6%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add cross-attention to the PickScore baseline | MHP test set, Overall · preference accuracy (%) | 70.65 -> 74.11 | Helped | Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · 23 citations) |
| disable preference-pair re-weighting (alpha 0.0) | VBench · Total | 81.93 -> 80.89 | Hurt | VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · 18 citations) |
| remove timestep conditioning from SPM | Pick-a-Pic val. · Aesthetic score | 5.887 -> 5.678 | Hurt | Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization (CVPR 2025 · 10 citations) |
| Add curriculum learning to LoRA + DPO | D1 · text alignment | 0.7502±0.0045 -> 0.7548±0.0041 | Helped | Curriculum Direct Preference Optimization for Diffusion and Consistency Models (CVPR 2025 · 8 citations) |
| replace the BLIP backbone with CLIP | ImageRewardDB, HPD v2, and Pick-a-Pic · preference prediction accuracy | 65.9 vs 66.3; 79.1 vs 79.4; 67.3 vs 67.1; Average 70.3 vs 70.5 | Mixed | Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · 8 citations) |
| Use KTO utility instead of Loss-Averse utility | HPS v2 test · PickScore | Helped | Aligning Diffusion Models by Optimizing Human Utility (NeurIPS 2024 · 6 citations) | |
| Add sigmoid loss weighting instead of constant weighting | SDXL multi-reward · MPS Win (%) | 56.5 -> 61.2 | Helped | Calibrated Multi-Preference Optimization for Aligning Diffusion Models (CVPR 2025 · 5 citations) |
| Use all three scoring dimensions: CLIP, VQA, and aesthetic | TIFA · VQA score | 77.1 -> 78.7 | Helped | AGFSync: Leveraging AI-Generated Feedback for Preference Optimization in Text-to-Image Generation (AAAI 2025 · 4 citations) |
| Use alpha-divergence with alpha=0.6 instead of Jensen-Shannon divergence | HPS-V2 test · LPIPS | 0.2989 ± 0.0361 vs 0.3109 ± 0.0373 | Helped | Generalizing Alignment Paradigm of Text-to-Image Generation with Preferences Through f-Divergence Minimization (AAAI 2025 · 4 citations) |
| train in one stage with real images as winners and generated images as losers | Human image generation test set · FID | 37.34 -> 112.67 | Hurt | Boost Your Human Image Generation Model via Direct Preference Optimization (CVPR 2025 · 3 citations) |
| interpolate user-embedding weights among CLIP, Aesthetic, and HPS objectives at inference | PartiPrompts · reward score | Mixed | Personalized Preference Fine-tuning of Diffusion Models (CVPR 2025 · 3 citations) | |
| replace BLIP backbone with CLIP | MVReward test set · accuracy | 83.1 -> 81.7 | Hurt | MVReward: Better Aligning and Evaluating Multi-View Diffusion Models with Human Preferences (AAAI 2025 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 32 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Pick-a-Pic v2 · 7 of 32 papers
- Parti-Prompts · 7 of 32 papers
- Pick-a-Pic · 5 of 32 papers
- HPDv2 · 4 of 32 papers
- HPS v2 · 2 of 32 papers
- GenEval · 2 of 32 papers
- HPSV2 benchmark · 2 of 32 papers
Baselines
- Diffusion-DPO · 9 of 32 papers
- DDPO · 6 of 32 papers
- SFT · 4 of 32 papers
- MAPO · 3 of 32 papers
- D3PO · 3 of 32 papers
- Diffusion-KTO · 3 of 32 papers
- VADER · 3 of 32 papers
- DPOK · 2 of 32 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 82 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 8 papers
- Zhejiang University · 6 papers
- Google · 3 papers
- Korea Advanced Institute of Science and Technology · 3 papers
- Georgia Institute of Technology · 2 papers
- Google DeepMind · 2 papers
- Hong Kong University of Science and Technology · 2 papers
- Meta · 2 papers
Use it
Download BibTeX of 23 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)
- T2I evaluation and controllable adapters (149 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).