Zoom-IQA-7B
Official model for Zoom-IQA: Image Quality Assessment with Reliable Region-Aware Reasoning.
Paper · Project page · Base model
Zoom-IQA-7B performs no-reference image quality assessment through two-round, region-aware reasoning. This specific version is our enhanced open-source release, trained on 8× NVIDIA H200 GPUs to achieve stronger generalization capabilities. This repository contains the inference weights; training data are released separately in the companion GR-IQA repository.
Quick start
Use the companion Zoom-IQA evaluator, which implements the prompts, crop routing, parsing, and evaluation metrics used by the model. In its Transformers environment, run:
zoomiqa-score \
--model-path /path/to/Zoom-IQA-7B \
--image /path/to/your_image.jpg \
--device 0
For benchmark evaluation, use zoomiqa-eval; a separate
zoomiqa-eval-vllm entrypoint is available for faster inference. See the code
repository for installation, annotation format, Python usage, and complete
command-line options.
The bundled generation_config.json is checkpoint metadata. The official
evaluator supplies its protocol settings explicitly.
Limitations
Outputs are sampled and may occasionally be malformed. Results can change with different prompts, decoding settings, preprocessing, or runtime versions. The model is intended for image-quality research, not high-stakes decisions.
License
Released under the Apache License 2.0. Benchmark images and GR-IQA source images retain their original licenses.
Citation
@article{liang2026zoomiqa,
title={Zoom-IQA: Image Quality Assessment with Reliable Region-Aware Reasoning},
author={Liang, Guoqiang and Wang, Jianyi and Wu, Zhonghua and Zhou, Shangchen and Loy, Chen Change},
journal={arXiv preprint arXiv:2601.02918},
year={2026}
}