kakaomacao/PR-IQA

PR-IQA: Partial-Reference Image Quality Assessment

Dense per-pixel quality assessment for generated images using partial 3D references.

Given a query (generated) image and a reference image, PR-IQA produces a dense per-pixel quality map by (1) generating a partial quality map via FeatureMetric (DINOv2 + VGGT + PyTorch3D) and (2) refining it into a dense quality map via a 3-input U-Net with cross-attention.

<p align="center"> <a href="https://kakaomacao.github.io/pr-iqa-project-page/"><img src="https://img.shields.io/badge/Project-Page-blue" alt="Project Page"></a> <a href="https://arxiv.org/abs/2604.04576"><img src="https://img.shields.io/badge/arXiv-2604.04576-b31b1b" alt="arXiv"></a> <a href="https://github.com/Kakaomacao/PR-IQA"><img src="https://img.shields.io/badge/GitHub-Code-black?logo=github" alt="GitHub"></a> </p>

Architecture

PR-IQA is a 3-input U-Net encoder-decoder with cross-attention (59.2M parameters).

• —Inputs: query image, reference image, and an internally generated partial quality map (+ 4-scale mask pyramid)
• —Encoder: 4 levels (48 → 96 → 192 → 384) with TransformerLikeBlocks (ChannelAttention + xformers Attention + FFN)
• —Decoder: 3 levels with skip connections from the query-image encoder
• —Output: sigmoid-activated per-pixel quality map in [0, 1]

Usage

Load from the Hub

from pr_iqa.model import PRIQA

model = PRIQA.from_pretrained("kakaomacao/PR-IQA")

Only the PR-IQA network weights are distributed here. To run the full pipeline (query + reference → dense quality map), clone the GitHub repository which includes the FeatureMetric stage and submodules (VGGT, LoftUp, PyTorch3D).

End-to-end inference

git clone --recursive https://github.com/Kakaomacao/PR-IQA.git
cd PR-IQA
pip install -e .

python inference.py \
    --checkpoint kakaomacao/PR-IQA \
    --generated examples/case1/query.png \
    --reference examples/case1/reference.jpg \
    --output output/quality_map.png

Citation

@article{priqa2026,
  title   = {PR-IQA: Partial-Reference Image Quality Assessment},
  author  = {Choi, Inseong and others},
  journal = {arXiv preprint arXiv:2604.04576},
  year    = {2026}
}

License

Apache License 2.0. Note that the full pipeline depends on submodules (VGGT, LoftUp) with their own licenses.

Acknowledgements

This research was supported by the MSIT (Ministry of Science and ICT), Korea, under the ITRC support program (IITP-2026-RS-2020-II201789), and the Artificial Intelligence Convergence Innovation Human Resources Development (IITP-2026-RS-2023-00254592) supervised by the IITP.