Research map / Image generation / Text-to-image diffusion
Learning Multi-Dimensional Human Preference for Text-to-Image Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds the MHP dataset of 918,315 human pairwise preferences over 607,541 text-to-image images across four dimensions and proposes MPS, a CLIP-based condition-masked cross-attention scorer that predicts preference under each dimension.
Problem
Existing human-preference scorers for text-to-image models collapse diverse human judgments into a single overall score, so they generalize poorly to individual dimensions such as aesthetics, semantic alignment, and detail quality.
Approach
A single unified scorer conditioned on a preference condition, implemented as a set of attribute words whose similarity to the prompt produces a binary mask that gates cross-attention between CLIP image and text tokens, so one model predicts multiple preference dimensions without retraining.
Key results
- MPS reaches 67.5% preference accuracy on ImageReward and 83.5% on HPD v2, above HPS v2 (65.7%, 83.3%).
- On MHP overall preference, MPS achieves 74.2% versus 69.5% for PickScore.
- On MHP multi-dimensional prediction, MPS scores 83.9 aesthetics, 83.9 alignment, 85.2 detail versus at most 74.8, 75.3, 62.1 for baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MPS instead of HPS v2 | text-to-image preference evaluation · ImageReward test set | preference accuracy (%): 65.7 -> 67.5 | Helped |
| Use MPS instead of HPS v2 | text-to-image preference evaluation · HPD v2 test set | preference accuracy (%): 83.3 -> 83.5 | Helped |
| Use MPS instead of PickScore | text-to-image preference evaluation · MHP test set, Overall | preference accuracy (%): 69.5 -> 74.2 | Helped |
| Use MPS instead of PickScore | multidimensional text-to-image preference evaluation · MHP test set, Overall | preference accuracy (%): 69.52 -> 74.24 | Helped |
| Use MPS instead of Aesthetic Score | multidimensional text-to-image preference evaluation · MHP test set, Aesthetics | preference accuracy (%): 82.85 -> 83.86 | Helped |
| Use MPS instead of CLIP score | multidimensional text-to-image preference evaluation · MHP test set, Alignment | preference accuracy (%): 82.69 -> 83.87 | Helped |
| Use MPS instead of HPS v2 | multidimensional text-to-image preference evaluation · MHP test set, Detail | preference accuracy (%): 62.06 -> 85.18 | Helped |
| Add cross-attention to the PickScore baseline (ablation) | MPS on MHP · MHP test set, Overall | preference accuracy (%): 70.65 -> 74.11 | Helped |
| Add cross-attention to the PickScore baseline (ablation) | MPS on MHP · MHP test set, Aesthetics | preference accuracy (%): 71.05 -> 71.83 | Helped |
| Add cross-attention to the PickScore baseline (ablation) | MPS on MHP · MHP test set, Alignment | preference accuracy (%): 70.81 -> 73.02 | Helped |
| Add cross-attention to the PickScore baseline (ablation) | MPS on MHP · MHP test set, Detail | preference accuracy (%): 57.46 -> 58.79 | Helped |
| Add the preference condition mask to cross-attention (ablation) | MPS on MHP · MHP test set, Aesthetics | preference accuracy (%): 71.83 -> 83.86 | Helped |
Surprising findings
- Preference models trained on a single overall score correlate poorly with detail quality, suggesting the preference dimensions are not strongly correlated.
- A separately trained model per preference underperforms the unified multi-condition model.
Limitations
- Preference accuracy for baselines is evaluated with public checkpoints, and the MHP overall numbers for HPS and HPS v2 are identical, indicating possible evaluation overlap or reporting issues.
- Condition word sets are hand-defined and the mask threshold is not extensively analyzed.
- Annotation relies on a crowdsourced team with limited description of demographic diversity.
Experimental setup
Datasets: MHP ImageReward test set HPD v2 test set DiffusionDB PromptHero
Benchmarks: ImageReward test set HPD v2 test set MHP test set
Metrics: preference accuracy
Baselines: CLIP score Aesthetic score ImageReward HPS HPS v2 PickScore
Builds on: CLIP PickScore HPS ImageReward
Closest papers by meaning
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.89)
- DreamAlign: Dynamic Text-to-3D Optimization with Human Preference Alignment (AAAI 2025 · similarity 0.89)
- MVReward: Better Aligning and Evaluating Multi-View Diffusion Models with Human Preferences (AAAI 2025 · similarity 0.88)
- Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms (NeurIPS 2024 · similarity 0.88)
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.87)
- MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation? (NeurIPS 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.