MR-IQA

MR-IQA: A Unified Margin View of Regression and Ranking for Blind Image Quality Assessment

<p align="center"> <a href="https://arxiv.org/abs/2606.29760">arXiv</a> | <a href="https://github.com/RobinY99/MR-IQA">GitHub</a> </p> <p align="center"> <a href="https://github.com/RobinY99/MR-IQA/blob/main/assets/mr_iqa_overview.pdf"> <img src="https://raw.githubusercontent.com/RobinY99/MR-IQA/main/assets/mr_iqa_overview.png" alt="MR-IQA unified margin view and training pipeline" width="96%"> </a> </p>

We derive that regression and ranking are approximately equivalent under a unified margin view. Based on this observation, we propose MR-IQA for margin learning in blind image quality assessment.

Validation Snapshot

The released checkpoint was validated after each epoch with an 8-shard setup on a held-out KONIQ split.

Best SRCC was reached at epoch 3. The final released checkpoint corresponds to epoch 10. Sanitized training metadata is available in training_guidance/.

Quick Start

Load the model with a standard Transformers vision-language workflow. The training and evaluation code use a no-reasoning prompt and parse the final numeric score from <answer>...</answer>.

System prompt:

You are an image quality assessment assistant. Output only the final score in <answer> </answer> tags.

User prompt:

What is your overall rating on the quality of this picture? The rating should be a float between 1 and 5, rounded to two decimal places, with 1 representing very poor quality and 5 representing excellent quality. Please only output the final answer with one score in <answer> </answer> tags.

Output Format

The expected response is only one score in answer tags:

<answer>3.74</answer>

The evaluation parser first reads the number inside <answer>...</answer> and clamps valid scores to the 1 to 5 range.

Citation

@misc{li2026mriqaunifiedmarginview,
  title={MR-IQA: A Unified Margin View of Regression and Ranking for Blind Image Quality Assessment},
  author={Yuan Li and Youyuan Lin and Zitang Sun and Yung-Hao Yang and Kiyofumi Miyoshi and Chenhui Chu and Shin'ya Nishida},
  year={2026},
  eprint={2606.29760},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.29760}
}