Research map / Image generation / Text-to-image diffusion
HPSv3: Towards Wide-Spectrum Human Preference Score
Yuhang Ma, Xiaoshi Wu, Keqiang Sun, Hongsheng Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Qwen2VL-7B backbone instead of Qwen2VL-2B (ablation) | HPSv3 · HPDv3 test set | preference accuracy (%): 66.3 -> 76.9; gain of 10.6 pp | Improved |
| use Qwen2VL-7B backbone instead of CLIP ViT-H/14 (ablation) | HPSv3 · HPDv3 test set | preference accuracy (%): 63.5 -> 76.9; gain of 13.4 pp | Improved |
| add uncertainty-aware ranking loss instead of ranknet loss (ablation) | HPSv3 · HPDv3 test set | preference accuracy (%): 76.3 -> 76.9 | Improved |
| add uncertainty-aware ranking loss instead of ranknet loss (ablation) | HPSv3 · PickScore test set | preference accuracy (%): 70.6 -> 72.8 | Improved |
| train on full HPDv3 data instead of HPDv2 data only (ablation) | HPSv3 · HPDv3 test set | preference accuracy (%): 66.6 -> 76.9 | Improved |
| train on full HPDv3 data instead of ImageReward data only (ablation) | HPSv3 · HPDv3 test set | preference accuracy (%): 63.4 -> 76.9 | Improved |
| train on full HPDv3 data instead of PickScore data only (ablation) | HPSv3 · HPDv3 test set | preference accuracy (%): 64.9 -> 76.9 | Improved |
| run more model-wise rounds in CoHP (1 to 4) (ablation) | CoHP · HPDv3 Benchmark (1000 prompts) | HPSv3 score: 11.34 -> 11.69 | Improved |
| run a fifth model-wise round in CoHP (ablation) | CoHP · HPDv3 Benchmark (1000 prompts) | HPSv3 score: 11.69 -> 11.65 | No clear effect |
| run more sample-wise rounds in CoHP (1 to 4) (ablation) | CoHP · HPDv3 Benchmark (1000 prompts) | HPSv3 score: 11.59 -> 12.84 | Improved |
| use HPSv3 instead of ImageReward as CoHP preference model | CoHP · user study on 100 text-image pairs | win rate: 87% | Improved |
| use HPSv3 as reward in DanceGRPO instead of HPSv2 | SD1.4 with DanceGRPO · qualitative | reward hacking | Improved |
Later papers that compared with this method
9 later accepted papers report HPSv3 as a baseline in their results tables; 7 of them report a better number on most of the results they share with it.
- Beyond VLM-Based Rewards: Diffusion-Native Latent Reward Modeling (ICML 2026 · better on 0 of 5 shared results)
- Arena-T2I Hard: Benchmarking and Improving Faithfulness with Dependency-Aware Checklist Rewards (NeurIPS 2026 · better on 4 of 4 shared results)
- PosterReward: Unlocking Accurate Evaluation for High-Quality Graphic Design Generation (CVPR 2026 · better on 4 of 4 shared results)
- Too Vivid to Be Real? Benchmarking and Calibrating Generative Color Fidelity (CVPR 2026 · better on 3 of 3 shared results)
- Consistent Noisy Latent Rewards for Trajectory Preference Optimization in Diffusion Models (ICLR 2026 · better on 1 of 2 shared results)
- TerraVis: Towards Evaluation of World-Grounded Visual Consistency in Text-to-Image Generation via MLLM Workflows (NeurIPS 2026 · better on 2 of 2 shared results)
- MaSC: A Masked Similarity Metric for Evaluating Concept-Driven Generation (NeurIPS 2026 · better on 1 of 1 shared result)
- StyleDoctor: Towards Specialist Reward Model for Style-centric Generation Tasks (CVPR 2026 · better on 1 of 1 shared result)
1 more not listed.
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ImageReward — ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation (NeurIPS 2023) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · similarity 0.90)
- DreamAlign: Dynamic Text-to-3D Optimization with Human Preference Alignment (AAAI 2025 · similarity 0.87)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.86)
- SafeSora: Towards Safety Alignment of Text2Video Generation via a Human Preference Dataset (NeurIPS 2024 · similarity 0.86)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.86)
- MVReward: Better Aligning and Evaluating Multi-View Diffusion Models with Human Preferences (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
