Research map / Image generation / Text-to-image diffusion
Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image Alignment
Ying Ba, Tianyu Zhang, Yalong Bai, Wenyi Mo, Tao Liang, Bing Su, Ji-Rong Wen
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 ICT reward instead of CLIP reward to optimize SD3.5-Turbo | SD3.5-Turbo · GenEval | Mean: 0.13 -> 0.71 | Improved |
| optimize SD3.5-Turbo with ICT reward | SD3.5-Turbo · GenEval | Mean: 0.69 -> 0.71 | Improved |
| optimize SD3.5-Turbo with PickScore reward | SD3.5-Turbo · GenEval | Mean: 0.69 -> 0.66 | Hurt |
| optimize SD3.5-Turbo with HP reward | SD3.5-Turbo · Parti-Prompts | Aesthetic Score: 6.293 -> 6.448 | Improved |
| optimize SD3.5-Turbo with HP reward | SD3.5-Turbo · Parti-Prompts | JPEG compressibility: 313.10 -> 334.86 | Improved |
| optimize SD3.5-Turbo with ICT reward | SD3.5-Turbo · Parti-Prompts | Aesthetic Score: 6.293 -> 6.285 | No clear effect |
| use CLIP as reward to optimize SD3.5-Turbo | SD3.5-Turbo · Parti-Prompts | Aesthetic Score: 6.293 -> 5.175 | Hurt |
| combine ICT and HP rewards multiplicatively instead of ICT alone (ablation) | SD3.5-Turbo · Parti-Prompts | Aesthetic Score: 6.285 -> 6.300 | Improved |
| use ICT-HP model instead of PickScore for preference prediction | reward model · Pick-High + Pickapic_v2 test | average accuracy: 79.04 -> 88.84 | Improved |
| add HP image-only stage on top of ICT model (ablation) | reward model · Pick-High + Pickapic_v2 test | average accuracy: 87.58 -> 88.47 | Improved |
| use ICT-HP model instead of PickScore on I2 > I1 pairs | reward model · Pick-High + Pickapic_v2 test | pairwise accuracy I2>I1: 74.80 -> 66.42 | Hurt |
| optimize SD3.5-Turbo with HPS_v2, PickScore, ImageReward rewards | SD3.5-Turbo · Table 5 multi-metric | multiple reward metrics | Mixed |
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 1 of 1 shared result
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
- Focus-N-Fix: Region-Aware Fine-Tuning for Text-to-Image Generation (CVPR 2025 · similarity 0.90)
- CHATS: Combining Human-Aligned Optimization and Test-Time Sampling for Text-to-Image Generation (ICML 2025 · similarity 0.89)
- Rich Human Feedback for Text-to-Image Generation (CVPR 2024 · similarity 0.89)
- MMIG-Bench: Towards Comprehensive and Explainable Evaluation of Multi-Modal Image Generation Models (NeurIPS 2025 · similarity 0.89)
- Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · similarity 0.89)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.89)
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.
