Research map / Image generation / Text-to-image diffusion
What Makes a Good Generated Image? Investigating Human and Multimodal LLM Image Preference Alignment
Rishab Parthasarathy, Jasmine Collins, Cory Stephenson
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 GPT-4o instead of Gemini-1.5-Flash as image judge | MLLM judge · pairwise preference, overall | Pearson vs human: 0.362 -> 0.461 | Improved |
| use finetuned PickScore instead of GPT-4o | image judge · pairwise preference, overall | Pearson vs human: 0.461 -> 0.498 | Improved |
| finetune LLaVA on task-specific synthetic data | LLaVA-LLaMA-3-8B · synthetic composition | accuracy: 52.0% -> 91.2% | Improved |
| finetune LLaVA on anatomy synthetic data | LLaVA-LLaMA-3-8B · synthetic anatomy | accuracy: 50.0% -> 57.6% | No clear effect |
| finetune LLaVA on style synthetic data | LLaVA-LLaMA-3-8B · synthetic style | accuracy: 57.2% -> 81.8% | Improved |
| finetune LLaVA on aesthetic synthetic data | LLaVA-LLaMA-3-8B · synthetic aesthetic | Pearson correlation: -0.010 -> 0.847 | Improved |
| prompt MLLMs to output joint angles before anatomy decision | GPT/Gemini/Claude · synthetic anatomy | accuracy: 46% and 54% | No clear effect |
| mix LLaVA baseline data with CaD data in training (ablation) | LLaVA multi-image · synthetic tasks | overfitting | Improved |
| finetune on a mix of tasks instead of single task (ablation) | LLaVA · synthetic tasks | accuracy | No clear effect |
Closest papers by meaning
- MMIG-Bench: Towards Comprehensive and Explainable Evaluation of Multi-Modal Image Generation Models (NeurIPS 2025 · similarity 0.90)
- MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation? (NeurIPS 2025 · similarity 0.89)
- Are High-Quality AI-Generated Images More Difficult for Models to Detect? (ICML 2025 · similarity 0.88)
- Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · similarity 0.88)
- Rich Human Feedback for Text-to-Image Generation (CVPR 2024 · similarity 0.87)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · 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.
