Research map / Image generation / Text-to-image diffusion
ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| train the BLIP-based ImageReward preference model rather than use BLIP Score | text-to-image human preference prediction · 466-prompt test set | preference accuracy: 57.76 -> 65.14 | Helped |
| train the BLIP-based ImageReward preference model rather than use Aesthetic Score | text-to-image human preference prediction · 466-prompt test set | preference accuracy: 57.35 -> 65.14 | Helped |
| train the BLIP-based ImageReward preference model rather than use CLIP Score | text-to-image human preference prediction · 466-prompt test set | preference accuracy: 54.82 -> 65.14 | Helped |
| use BLIP instead of CLIP as the reward-model backbone (ablation) | ImageReward · 4k training set | preference accuracy: 61.87 -> 64.71 | Helped |
| use BLIP instead of CLIP as the reward-model backbone (ablation) | ImageReward · 8k training set | preference accuracy: 62.98 -> 65.14 | Helped |
| increase the BLIP training set from 1k to 8k (ablation) | ImageReward · human preference prediction | preference accuracy: 63.07 -> 65.14 | Helped |
| increase the BLIP training set from 2k to 8k (ablation) | ImageReward · human preference prediction | preference accuracy: 63.18 -> 65.14 | Helped |
| increase the BLIP training set from 4k to 8k (ablation) | ImageReward · human preference prediction | preference accuracy: 64.71 -> 65.14 | Helped |
| increase the CLIP training set from 4k to 8k (ablation) | ImageReward · human preference prediction | preference accuracy: 61.87 -> 62.98 | Helped |
| fine-tune diffusion models directly using reward gradients at later denoising steps | Stable Diffusion v1.4 · real-user and challenging prompts | human preference | Helped |
| use RAFT fine-tuning instead of the baseline | Stable Diffusion v1.4 · real-user and challenging prompts | human preference | No clear effect |
| use Reward Weighted fine-tuning instead of the baseline | Stable Diffusion v1.4 · real-user and challenging prompts | human preference | No clear effect |
Closest papers by meaning
- Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · similarity 0.89)
- Subject-driven Text-to-Image Generation via Preference-based Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Rich Human Feedback 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)
- Automated Filtering of Human Feedback Data for Aligning Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.89)
- Aligning Diffusion Models by Optimizing Human Utility (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.