Research map / Vision-language and multimodal / Multimodal LLMs
MJ-Bench: Is Your Multimodal Reward Model Really a Good Judge for Text-to-Image Generation?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MJ-Bench, a preference benchmark for evaluating multimodal judges (CLIP-style score models, open and closed VLMs) as feedback providers for text-to-image generation across alignment, safety, quality, and bias, and validates judges via human ranking of RLAIF-finetuned SD-1.5 models.
Problem
Multimodal judges (reward models) used to align text-to-image models are inadequately evaluated for accuracy, safety, quality, and bias, risking misaligned fine-tuning.
Approach
A triplet-based preference benchmark decomposing four perspectives into verifiable subcategories, scoring judges by pairwise accuracy (and bias metrics GES/NDS), plus end-to-end validation by fine-tuning SD-1.5 with each judge's feedback and collecting human rankings.
Key results
- GPT-4o achieves the best average performance across perspectives among judges; in safety it reaches 100.0% accuracy without ties and 97.6% on artifacts.
- Human ranking of SD-1.5 models fine-tuned via DPO shows GPT-4o achieving the best alignment (AR 2.50, AV 17.21%) and bias (AR 2.10, AV 21.58%) performance.
- Open-source VLMs give more accurate feedback on Likert scales than numerical scales, while closed-source VLMs are more consistent across scales.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GPT-4o as the multimodal judge instead of other evaluated judges | MJ-Bench judges · MJ-Bench overall | average accuracy | Helped |
| use CLIP-based scoring models rather than open-source VLMs | multimodal judges · MJ-Bench text-image alignment and artifact | accuracy | Helped |
| use VLM judges rather than CLIP-based scoring models | multimodal judges · MJ-Bench safety and bias | accuracy | Helped |
| fine-tune SD-1.5 using GPT-4o feedback | SD-1.5 · Alignment, safety, and bias human evaluation | ranking and voting: 2.50; 1.89; 2.10; 17.21%; 17.37%; 21.58% | Helped |
| use Gemini Ultra feedback rather than other judges | SD-1.5 safety fine-tuning · Safety human evaluation | average ranking: 1.64 | Helped |
| input both images together rather than separately (ablation) | open-source multi-input VLM judges · MJ-Bench alignment, safety, and artifact | accuracy | Mixed |
| prioritize the chosen image in pairwise input (ablation) | InternVL-chat-v1-5 · MJ-Bench alignment | accuracy: 32.8 -> 75.8 | Helped |
| prioritize the rejected image in pairwise input (ablation) | Qwen-VL-Chat · MJ-Bench alignment | accuracy: 29.1 -> 73.0 | Helped |
| use a 5-point Likert scale rather than a numerical rating (ablation) | open-source VLM judges · MJ-Bench alignment | accuracy | Helped |
| use numerical scale [0, 10] rather than other scales (ablation) | VLM judges overall · MJ-Bench alignment | accuracy: 37.6 | Helped |
| use DPO rather than DDPO for fine-tuning (ablation) | SD-1.5 using multimodal judge feedback · Human evaluation across alignment, safety, and bias | ranking stability | Helped |
| fine-tune with Claude 3 Opus feedback rather than GPT-4o or GPT-4-vision feedback | SD-1.5 · Text-image alignment | human evaluation | Hurt |
Surprising findings
- Open-source multi-input VLMs (Qwen-VL-Chat, InternVL-chat-v1.5) show strong order inconsistency, with preferences flipping depending on which image is shown first.
- Smaller score models such as HPS-v2.1 outperform some open-source VLMs and even some closed-source VLMs on alignment and bias in human evaluation.
Limitations
- Safety and some subcategories have small sample sizes (e.g., 92 toxicity and 472 NSFW pairs).
- Human evaluation covers only six judges with 100 prompts per perspective, limiting statistical power.
- Data generated with SDXL and inpainting may not fully reflect real-world distributions.
Experimental setup
Datasets: Pick-a-pic HPDv2 ImageRewardDB I2P NudeNet-detected NSFW images MPII
Benchmarks: MJ-Bench
Metrics: accuracy GES NDS average ranking voting
Baselines: CLIP-v1 BLIP-v2 PickScore-v1 HPS-v2.1 ImageReward Aesthetics LLaVA-1.5 InternVL-chat-v1.5 Qwen-VL-Chat Idefics2-8b GPT-4-vision GPT-4o
Builds on: DPO DDPO Pick-a-pic HPDv2 ImageRewardDB
Closest papers by meaning
- MMIG-Bench: Towards Comprehensive and Explainable Evaluation of Multi-Modal Image Generation Models (NeurIPS 2025 · similarity 0.90)
- OpenING: A Comprehensive Benchmark for Judging Open-ended Interleaved Image-Text Generation (CVPR 2025 · similarity 0.90)
- OneIG-Bench: Omni-dimensional Nuanced Evaluation for Image Generation (NeurIPS 2025 · similarity 0.89)
- Rich Human Feedback for Text-to-Image Generation (CVPR 2024 · similarity 0.89)
- SILMM: Self-Improving Large Multimodal Models for Compositional Text-to-Image Generation (CVPR 2025 · similarity 0.89)
- MEGA-Bench: Scaling Multimodal Evaluation to over 500 Real-World Tasks (ICLR 2025 · similarity 0.88)
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.