Research map / Image generation / Text-to-image diffusion
Aligning Multi-Character Narrative Image Generation with Multi-Aspect Human Preferences
Ziyi Gao, Zhipeng Wei, Jingjing Chen, Zhiyu Tan, Hao Li, Yi-Ping Phoebe Chen
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 |
|---|---|---|---|
| add critique generation before scoring in the reward model (ablation) | NIReward · NI-Bench visual quality | preference accuracy (%): 72.66 -> 83.00 | Improved |
| add critique generation before scoring (ablation) | NIReward · NI-Bench prompt following | preference accuracy (%): 83.83 -> 86.07 | Improved |
| add critique generation before scoring (ablation) | NIReward · NI-Bench identity consistency | preference accuracy (%): 84.60 -> 85.10 | Improved |
| train NIReward instead of using GPT-4o-mini as judge | NIReward · NI-Bench identity consistency | preference accuracy (%): 46.75 -> 85.10 | Improved |
| use NIReward instead of GPT-4o-mini on visual quality | NIReward · NI-Bench visual quality | preference accuracy (%): 88.42 -> 83.00 | Hurt |
| add rejection sampling to ADPO (ablation) | ADPO · Preference optimization | ImageReward / NIReward P.F.: w/o RS 0.951 -> 1.010 ImageReward; 0.001 -> 0.079 P.F. | Improved |
| add adaptive weighted learning to ADPO (ablation) | ADPO · Preference optimization | NIReward identity consistency: -0.013 -> 0.010 | Improved |
| remove adaptive weighting (w/o AW) (ablation) | ADPO · Preference optimization | NIReward prompt following: 0.079 (full) vs 0.123 (w/o AW) | Mixed |
| use Dominating Comparison instead of mean scores (ablation) | ADPO · Preference optimization | ID Sim: 26.94 -> 28.88 | Improved |
| use ADPO with NIReward instead of Diffusion-DPO baselines | PhotoMaker V2 · Preference optimization | NIReward identity consistency: -0.036 -> 0.010 | Improved |
| optimize with ADPO with NIReward | PhotoMaker V2 · Preference optimization | CLIP: 32.38 -> 32.80 | Improved |
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
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.91)
- Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · similarity 0.89)
- DreamAlign: Dynamic Text-to-3D Optimization with Human Preference Alignment (AAAI 2025 · similarity 0.88)
- Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.87)
- MVReward: Better Aligning and Evaluating Multi-View Diffusion Models with Human Preferences (AAAI 2025 · similarity 0.87)
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.
