Research map / Image generation / Text-to-image diffusion
Multimodal LLMs as Customized Reward Models for Text-to-Image Generation
Shijie Zhou, Ruiyi Zhang, Huaisheng Zhu, Branislav Kveton, Yufan Zhou, Jiuxiang Gu, Jian Chen, Changyou 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 SkipCA reward head (vs MLP) (ablation) | LLaVA-Reward-Phi · MJ-Bench safety | accuracy w/ tie: 39.7 -> 55.2 | Improved |
| add SkipCA reward head (vs MLP) (ablation) | LLaVA-Reward-Phi · MJ-Bench artifact (fidelity) | accuracy w/ tie: 87.3 -> 91.1 | Improved |
| add SkipCA reward head (vs MLP) (ablation) | LLaVA-Reward-Phi · MJ-Bench alignment | accuracy w/ tie: 68.2 -> 66.1 | Hurt |
| use GPM objective instead of BT (ablation) | LLaVA-Reward-Phi · MJ-Bench alignment | accuracy w/ tie: 65.2 -> 66.1 | Improved |
| add 70k hard-negative preference pairs to training (ablation) | LLaVA-Reward-Phi · MJ-Bench alignment | accuracy w/ tie: 64.5 -> 66.1 | Improved |
| use EOS-token hidden state instead of mean over tokens (ablation) | LLaVA-Reward-Phi · MJ-Bench alignment | accuracy w/ tie: 54.6 -> 66.1 | Improved |
| use LoRA instead of full fine-tuning (ablation) | LLaVA-Reward-Phi · MJ-Bench artifact (fidelity) | accuracy w/ tie: 86.3 -> 91.1 | Improved |
| use Qwen2.5-VL backbone instead of Phi-3.5-vision (ablation) | LLaVA-Reward · MJ-Bench artifact (fidelity) | accuracy w/ tie: 91.1 -> 94.3 | Improved |
| use LLaVA-NeXT 13.4B instead of Phi-3.5-vision (ablation) | LLaVA-Reward · MJ-Bench artifact (fidelity) | accuracy w/ tie: 91.1 -> 90.8 | No clear effect |
| use middle layer (21) hidden state instead of last layer (32) (ablation) | LLaVA-Reward · SMID safety | F1 score | Hurt |
| use visual tokens after projector in SkipCA rather than deeper layers (ablation) | LLaVA-Reward · MJ-Bench safety | accuracy | Improved |
| steer SDXL with LLaVA-Reward in FK steering | SDXL · GenEval | overall score: 0.563 -> 0.645 | 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 3 of 3 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
- LMFusion: Adapting Pretrained Language Models for Multimodal Generation (NeurIPS 2025 · similarity 0.90)
- SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation (CVPR 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)
- A Comprehensive Study of Decoder-Only LLMs for Text-to-Image Generation (CVPR 2025 · similarity 0.89)
- Learning to Generate via Understanding: Understanding-Driven Intrinsic Rewarding for Unified Multimodal Models (CVPR 2026 · similarity 0.89)
- MM-RLHF: The Next Step Forward in Multimodal LLM Alignment (ICML 2025 · 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.
