Qwen-Image-Bench

Qwen-Image-Bench

A judge model called Q-Judger that grades AI-made images against the prompt that asked for them, scoring five groups of criteria and replying in JSON. For people who test text-to-image models.

Made by Qwen on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen-Image-Bench is the judge model of a benchmark for text-to-image generation, and its card calls it Q-Judger. It is a 27.4B vision-language model fine-tuned from Qwen3.6-27B. You give it the prompt and the picture a generator made from it. It works through a checklist of criteria, reasons step by step, then returns a JSON result with a score for each criterion.

Who it suits: Researchers and engineers who compare or tune image generators and want a model to do the grading instead of people.

Industry | Who uses it | How they use it | Example
AI research | Image model teams | Score outputs of a new generator across many prompts | Compare two checkpoints on the same prompt set
Product teams | ML engineers | Add an automatic quality gate to an image pipeline | Flag pictures that fail a criterion
Benchmarking | Evaluation researchers | Run the Qwen-Image-Bench benchmark | Grade a model's pictures on all five groups

Anyone who just wants to make pictures can skip it, since this model grades them and does not draw.

Where to use it:

• The transformers library in Python, with the Safetensors weights.
• The Qwen-Image-Bench project on GitHub and its benchmark dataset, which the card links.
• Windows, macOS or Linux: the download steps run on all three with Python.

Use case | Who | What they get back
Image grading | Researchers | A JSON score for each criterion
Model comparison | ML engineers | Scores that line up for two generators on one prompt
Quality gates | Pipeline builders | A pass or fail signal per picture

How to set it up: What you need first:

• Python with pip.
• The transformers library.
• A GPU with room for the weights: about 54.7 GB in BF16, or about 13.7 GB at 4-bit, as estimated on the record.

1. Install the Hugging Face hub package with pip.
2. Download the repository with the hf command.
3. Load it with transformers, following the model page.
4. Pass in a prompt and an image, and read the JSON that comes back.

pip install -U huggingface_hub
hf download Qwen/Qwen-Image-Bench

How to use it: Send the model the text prompt plus the generated image. It thinks first, then writes the scores as JSON. The text available here leaves out the exact calling code, so use the model page and the GitHub project for it.

Title: Parts of the benchmark
Tool | What it does
Q-Judger model | Grades a picture against its prompt
Benchmark dataset | The set of prompts to test generators with
GitHub project | The code that runs the evaluation
transformers | Loads the judge model in Python

Example: You feed in the prompt "a red bicycle leaning on a brick wall" and the picture a generator made from it. The model checks each criterion and replies with a JSON object where every one is marked 0, 1, 2 or N/A.

For newcomers and experts: People used to judge generated pictures by eye, which is slow. A judge model does the same job quickly and the same way each time. A good first step is to grade a few pictures you already have and read the JSON.

The criteria sit in a three-level tree under five top-level groups. Under Quality, for example, Realism covers physical logic and material texture, and Detail covers noise, edge clarity and naturalness. The record lists English and Chinese as its languages.

Take a prompt and its generated image
The judge reasons through each criterion
It writes JSON scores
You compare scores across models

Good to know: The record estimates how much space the weights take.

Title: Weights size, estimate
BF16: 54.7 GB
4-bit: 13.7 GB

• Scores: 0 means fail, 1 means pass, 2 means excellent, and N/A marks a criterion that does not apply.
• Size: about 27.4B parameters.
• Versions: the record lists 7 quantized versions made by others.
• License: apache-2.0
• Cost to run: no API key, account or paid service named.
• Data it reaches: files you download from Hugging Face, and the prompts and images you pass in.
• Not for: making images, or anyone without a large GPU for the full weights.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 27.4B
• Weights: about 54.7 GB in BF16; about 13.7 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 20,875
• Likes: 116
• License: apache-2.0
• Languages: en, zh
• Based on: Qwen/Qwen3.6-27B
• Quantized versions: 7
• Tags: transformers, safetensors, qwen3_5, image-text-to-text, judge-model, text-to-image, evaluation, benchmark, qwen, conversational

Links

• Model: https://huggingface.co/Qwen/Qwen-Image-Bench
• Files: https://huggingface.co/Qwen/Qwen-Image-Bench/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Qwen/Qwen-Image-Bench

Then follow the usage in its model card. Nothing in this kit runs by itself.