ReLIQS

Model weights for Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency (CVPR 2026).

Hakan Emre Gedik, Shashank Gupta, Alan Bovik
The University of Texas at Austin · University of Colorado Boulder

ReLIQS predicts image quality without a reference image using multiscale patches, CLIP features, and learned quality-aware saliency.

This repository hosts model weights only. Installation, inference, saliency visualization, and training instructions are available in the GitHub repository.

Usage

Download a compatible checkpoint and save it as model_weights.pth. After following the GitHub installation instructions, run from the code repository root:

python score_image.py \
    --checkpoint_path model_weights.pth \
    --image_path /path/to/image.jpg

The script uses EMA weights and returns a score in [0, 1], where higher is better. Scores are not calibrated to a dataset's original MOS scale.

Citation

If you use ReLIQS, its pretrained weights, or its code in your research, please cite our CVPR 2026 paper:

@InProceedings{Gedik_2026_CVPR,
    author    = {Gedik, Hakan Emre and Gupta, Shashank and Bovik, Alan},
    title     = {Learning Where to Look and How to Judge: Resolution-agnostic Image Quality Assessment with Quality-aware Saliency},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {37507-37517}
}