MJ-Bench
Models compare image pairs for text alignment, safety, visual quality, and bias. Scores measure preference accuracy, with separate dispersion and equality measures for bias.
Accuracy is reported with and without ties. Dataset access requires a request that the makers describe as automatically approved.
- Made by
- MJ-Bench team at UNC-Chapel Hill, University of Chicago, Stanford University, and collaborators
- Released
- July 2024
- Kind
- Image generation and editing
- Level
- Everyday
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- All public
- Measures
- Capability
- Risk
- Hate
- Language
- English
- Interaction
- Single turn
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Vision, Content safety
- Strongest model tested
- Claude 3 Opus #173 on the capability index
- Last model added
- April 2024
- Scores refreshed
- November 18th, 2026
In the authors’ words
While text-to-image models like DALLE-3 and Stable Diffusion are rapidly proliferating, they often encounter challenges such as hallucination, bias, and the production of unsafe, low-quality output. To effectively address these issues, it is crucial to align these models with desired behaviors based on feedback from a multimodal judge. Despite their significance, current multimodal judges frequently undergo inadequate evaluation of their capabilities and limitations, potentially leading to misalignment and unsafe fine-tuning outcomes. To address this issue, we introduce MJ-Bench, a novel benchmark which incorporates a comprehensive preference dataset to evaluate multimodal judges in providing feedback for image generation models across four key perspectives: alignment, safety, image quality, and bias. Specifically, we evaluate a large variety of multimodal judges including smaller-sized CLIP-based scoring models, open-source VLMs (e.g. LLaVA family), and close-source VLMs (e.g. GPT-4o, Claude 3) on each decomposed subcategory of our preference dataset. Experiments reveal that close-source VLMs generally provide better feedback, with GPT-4o outperforming other judges in average. Compared with open-source VLMs, smaller-sized scoring models can provide better feedback regarding text-image alignment and image quality, while VLMs provide more accurate feedback regarding safety and generation bias due to their stronger reasoning capabilities. Further studies in feedback scale reveal that VLM judges can generally provide more accurate and stable feedback in natural language (Likert-scale) than numerical scales. Notably, human evaluations on end-to-end fine-tuned models using separate feedback from these multimodal judges provide similar conclusions, further confirming the effectiveness of MJ-Bench. All data, code, models are available at https://huggingface.co/MJ-Bench.
GPT-4o (snapshot unspecified) tops the leaderboard with 70.9% InternVL-Chat-V1-5 leads with 44.5% among models released by April 2024
Best score per entry from the benchmark's own leaderboard, read on November 18th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. The best model we track is InternVL-Chat-V1-5, at 44.5%. This view shows only models we track, each at its best.
The newest model with a score here came out in April 2024, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 70.9 GPT-4o (snapshot unspecified)
- 70.0 GPT-4-vision
- 59.0 PickScore-v1
- 50.8 HPS-v2.1
- 50.7 AestheticsPredictor
- 50.2 Gemini Ultra
- 48.8 ImageReward
- 44.5 InternVL-Chat-V1-5
- 40.7 llava-v1.6-mistral-7b
- 40.0 Qwen-VL-Chat
- 36.8 llava-v1.6-vicuna-13b
- 35.8
Idefics2-8B
- 35.3 CLIP-v2
- 32.9 MiniGPT4-v2
- 31.6 Claude 3 Opus
- 30.2 LLaVA-1.5-7b
- 30.0
LLaVA-v1.5-13B
- 28.5 instructblip-vicuna-7b
- 28.1 BLIP-v2
- 23.9 Prometheus-Vision-7b
- 19.0 Prometheus-Vision-13b
- 1 GPT-4o (snapshot unspecified) Multi Image 70.9%
- 2 GPT-4-vision Multi Image 70%
- 3 PickScore-v1 Single Image 59%
- 4 HPS-v2.1 Single Image 50.8%
- 5 AestheticsPredictor Single Image 50.7%
- 6 Gemini Ultra Multi Image 50.2%
- 7 ImageReward Single Image 48.8%
- 8 InternVL-Chat-V1-5 Multi Image 44.5%
- 9 llava-v1.6-mistral-7b Single Image 40.7%
- 10 Qwen-VL-Chat Multi Image 40%
- 11 llava-v1.6-vicuna-13b Single Image 36.8%
- 12
Idefics2-8B Multi Image 35.8%
- 13 CLIP-v2 Single Image 35.3%
- 14 MiniGPT4-v2 Single Image 32.9%
- 15 Claude 3 Opus Multi Image 31.6%
- 16 LLaVA-1.5-7b Single Image 30.2%
- 17
LLaVA-v1.5-13B Single Image 30%
- 18 instructblip-vicuna-7b Single Image 28.5%
- 19 BLIP-v2 Single Image 28.1%
- 20 Prometheus-Vision-7b Single Image 23.9%
- 21 Prometheus-Vision-13b Single Image 19%
Source: Official MJ-Bench leaderboard
- 1 InternVL-Chat-V1-5 Multi Image 44.5%
- 2 llava-v1.6-mistral-7b Single Image 40.7%
- 3 Qwen-VL-Chat Multi Image 40%
- 4 llava-v1.6-vicuna-13b Single Image 36.8%
- 5
Idefics2-8B Multi Image 35.8%
- 6 Claude 3 Opus Multi Image 31.6%
- 7
LLaVA-v1.5-13B Single Image 30%
- 8 instructblip-vicuna-7b Single Image 28.5%
Source: Official MJ-Bench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | GPT-4o (snapshot unspecified) | 70.9% | Multi Image | Official leaderboard | |
| 2 | GPT-4-vision | 70% | Multi Image | Official leaderboard | |
| 3 | PickScore-v1 | 59% | Single Image | Official leaderboard | |
| 4 | HPS-v2.1 | 50.8% | Single Image | Official leaderboard | |
| 5 | AestheticsPredictor | 50.7% | Single Image | Official leaderboard | |
| 6 | Gemini Ultra | 50.2% | Multi Image | Official leaderboard | |
| 7 | ImageReward | 48.8% | Single Image | Official leaderboard | |
| 8 | InternVL-Chat-V1-5 | 44.5% | Multi Image | Apr 2024 | Official leaderboard |
| 9 | llava-v1.6-mistral-7b | 40.7% | Single Image | Jan 2024 | Official leaderboard |
| 10 | Qwen-VL-Chat | 40% | Multi Image | Aug 2023 | Official leaderboard |
| 11 | llava-v1.6-vicuna-13b | 36.8% | Single Image | Jan 2024 | Official leaderboard |
| 12 | | 35.8% | Multi Image | Apr 2024 | Official leaderboard |
| 13 | CLIP-v2 | 35.3% | Single Image | Official leaderboard | |
| 14 | MiniGPT4-v2 | 32.9% | Single Image | Official leaderboard | |
| 15 | Claude 3 Opus | 31.6% | Multi Image | Feb 2024 | Official leaderboard |
| 16 | LLaVA-1.5-7b | 30.2% | Single Image | Official leaderboard | |
| 17 | | 30% | Single Image | Oct 2023 | Official leaderboard |
| 18 | instructblip-vicuna-7b | 28.5% | Single Image | May 2023 | Official leaderboard |
| 19 | BLIP-v2 | 28.1% | Single Image | Official leaderboard | |
| 20 | Prometheus-Vision-7b | 23.9% | Single Image | Official leaderboard | |
| 21 | Prometheus-Vision-13b | 19% | Single Image | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | InternVL-Chat-V1-5 | 44.5% | Multi Image | Apr 2024 | Official leaderboard |
| 2 | llava-v1.6-mistral-7b | 40.7% | Single Image | Jan 2024 | Official leaderboard |
| 3 | Qwen-VL-Chat | 40% | Multi Image | Aug 2023 | Official leaderboard |
| 4 | llava-v1.6-vicuna-13b | 36.8% | Single Image | Jan 2024 | Official leaderboard |
| 5 | | 35.8% | Multi Image | Apr 2024 | Official leaderboard |
| 6 | Claude 3 Opus | 31.6% | Multi Image | Feb 2024 | Official leaderboard |
| 7 | | 30% | Single Image | Oct 2023 | Official leaderboard |
| 8 | instructblip-vicuna-7b | 28.5% | Single Image | May 2023 | Official leaderboard |
Rankings by part
We read 4 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as Alignment, Safety and Quality. Claude 3 Opus leads 1 of them, and 3 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Alignment
8 models we track have a score here.
- 1 Claude 3 Opus ↑5 57.1%
- 2 InternVL-Chat-V1-5 ↓1 55.3%
- 3 Qwen-VL-Chat 52.1%
- 4
Idefics2-8B ↑1 32.6%
- 5 llava-v1.6-mistral-7b ↓3 31.3%
- 6 llava-v1.6-vicuna-13b ↓2 29.1%
- 7 instructblip-vicuna-7b ↑1 17.1%
- 8
LLaVA-v1.5-13B ↓1 10.3%
Safety
8 models we track have a score here.
- 1
LLaVA-v1.5-13B ↑6 30.7%
- 2 llava-v1.6-vicuna-13b ↑2 27.9%
- 3 Qwen-VL-Chat 26.8%
- 4 instructblip-vicuna-7b ↑4 26.4%
- 5 llava-v1.6-mistral-7b ↓3 15.2%
- 6
Idefics2-8B ↓1 13.6%
- 7 Claude 3 Opus ↓1 13.4%
- 8 InternVL-Chat-V1-5 ↓7 6.30%
Quality
8 models we track have a score here.
- 1 InternVL-Chat-V1-5 66.3%
- 2
Idefics2-8B ↑3 46.1%
- 3 llava-v1.6-mistral-7b ↓1 45.8%
- 4 llava-v1.6-vicuna-13b 36.8%
- 5 instructblip-vicuna-7b ↑3 25.2%
- 6 Qwen-VL-Chat ↓3 23.6%
- 7
LLaVA-v1.5-13B 23.3%
- 8 Claude 3 Opus ↓2 11.9%
Bias
The official leaderboard combines this perspective score into its weighted overall score. 8 models we track have a score here.
- 1 Qwen-VL-Chat ↑2 71.9%
- 2 llava-v1.6-mistral-7b 69.9%
- 3
LLaVA-v1.5-13B ↑4 69.7%
- 4 Claude 3 Opus ↑2 57.7%
- 5 llava-v1.6-vicuna-13b ↓1 56.3%
- 6 instructblip-vicuna-7b ↑2 53.1%
- 7
Idefics2-8B ↓2 42.1%
- 8 InternVL-Chat-V1-5 ↓7 25.4%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official MJ-Bench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official MJ-Bench leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in Vision
- MMMU In use 11.5K college-level multimodal questions across 30 subjects. Multiple-choice and open accuracy.
Qwen 3.6 Plus 86%
- MMMU-Pro In use Harder MMMU: college-level multimodal questions with more options and vision-only inputs to remove text shortcuts. GPT-5.6 Sol 84.6%
- CharXiv In use Chart understanding on real arXiv figures. Reasoning questions require synthesizing information from complex charts. Claude Fable 5 93.5%
- MathVista In use Math reasoning in visual contexts (charts, geometry, puzzles) from 31 datasets. Answer accuracy.
Seed2.1 Pro 90.7%
- MATH-Vision (MathVision) In use 3,040 competition math problems with visual context. Answer accuracy.
Qwen 3.8 Max 97.7%
- MMStar In use Models answer 1,500 human-curated image questions that cannot be solved without the image, across 6 core capabilities. Scored by accuracy.
Qwen 3.8 Max 85.9%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 2 years ago, in April 2024, for InternVL-Chat-V1-5. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the official leaderboard (21 entries).
- Pass Models scored 8 models we track, plus 13 other entries on its leaderboard.
- Warning Newest score The newest came 2 years ago, in April 2024, for InternVL-Chat-V1-5.
- Note Frontier models None of the top 10 has a score. The strongest tested is Claude 3 Opus (#173).
- Pass Headroom The best is 44.5%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on huggingface.co: 8 rows for models we track and 13 other entries at the last read.
- Pass Last read November 18th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.