---
license: mit
library_name: transformers
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3-VL-2B-Instruct
tags:
- image-quality-assessment
- iqa
- qwen3-vl
- mr-iqa
---

# MR-IQA

MR-IQA is a 2B vision-language model for no-reference image quality assessment. It predicts a perceptual quality score from an input image and a short instruction prompt.

This repository contains the final checkpoint from the original June 17, 2026 MR-IQA 2B training run. The training data used KONIQ-style image quality supervision, and the model was validated on a 200-image KONIQ validation split with fixed data sampling.

## Validation

Final checkpoint validation on the 200-image KONIQ validation split:

| Split | Valid samples | SRCC | PLCC | Shards |
| --- | ---: | ---: | ---: | ---: |
| KONIQ validation | 200 | 0.9250992954 | 0.9420850686 | 8 |

## Files

The repository includes only inference-ready assets:

- 
- 
- 
- 
- 
- 
- 

Training-only artifacts such as optimizer state, scheduler state, RNG state, trainer state, and distributed checkpoint helpers are intentionally excluded.

## Usage

Use a standard Transformers vision-language loading flow with  and the Qwen3-VL model class. Provide an image and ask the model to assess overall perceptual quality. The expected score range is 1 to 5, rounded to two decimal places.

Example prompt:



The model is released as MR-IQA. Paths used during training and evaluation are not required for loading or inference.