RQI (Relative Quality Index)

Official code for the CVPR 2026 paper:

Bridging the Perception Gap in Image Super-Resolution Evaluation

📄 arXiv  |  🌐 Project Page

---

We investigate whether long-standing image quality metrics can actually evaluate modern SR models, and introduce RQI — a learned metric that better aligns with human preference for SR model evaluation.

---

Notice: RQI is trained on general distorted image data (the PIPAL dataset), making it potentially applicable beyond SR. We encourage exploration of its useages in other image quality assessment tasks.

1. Quick Use

Install via pip:

pip install rqi-iqa

You can then use RQI by a few lines:

from rqi import RQI

model = RQI(pretrained=True)

# score = model(test_image, gt_image)
score = model("imgs/0801_BSRGAN.png", "imgs/0801_GT.png")

RQI accepts various input formats: str (image path), PIL.Image, np.ndarray, or torch.Tensor

⚠️ RQI is assymetic: the first argument is the image to be evaluated; the second argument is the reference (ground truth) image.

Output: A score in [0, 1] — higher means better image quality.

---

2. For Development

2.1 Setup

Requirements: Python >= 3.8

git clone https://github.com/CVC-Color/rqi
cd rqi
pip install -r requirements.txt

2.2 Benchmark Evaluation

We provide a script for testing RQI on the SRIQA-Bench dataset:

python scripts/test_SRIQA.py

This script reports:

• Mean SRCC / PLCC over single images
• Overall SRCC / PLCC across the dataset

Performance

Our provided pretrained weights achieve better performance than reported in the paper.

Metric	Paper	Provided Weights
SRCC	0.733	0.780
PLCC	0.739	0.795
SRCC (subset)	0.609	0.590
PLCC (subset)	0.564	0.565

2.3 Using RQI as a Training Loss

🚧 TODO

---

Citation

If you find this work useful, please consider citing:

@inproceedings{su2026rqi,
  title     = {Bridging the Perception Gap in Image Super-Resolution Evaluation},
  author    = {Su, Shaolin and Rocafort, Josep M. and Xue, Danna and
               Serrano-Lozano, David and Sun, Lei and Vazquez-Corral, Javier},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision
               and Pattern Recognition (CVPR)},
  year      = {2026}
}

---

Acknowledgement

This work was supported by grants PID2021-128178OB-I00 and PID2024-162555OB-I00 funded by MCIN/AEI/10.13039/501100011033 and ERDF "A way of making Europe", the Generalitat de Catalunya CERCA Program, the grant Càtedra ENIA UAB-Cruïlla (TSI-100929-2023-2), and the 2025 Leonardo Grant for Scientific Research and Cultural Creation from the BBVA Foundation. Shaolin Su was supported by the HORIZON MSCA Postdoctoral Fellowship (project number 101152858). David Serrano-Lozano was supported by the FPI grant from the Spanish Ministry of Science and Innovation (PRE2022-101525). Lei Sun was partially funded by the Ministry of Education and Science of Bulgaria's support for INSAIT and by the European Union's Horizon Europe programme (grant agreement 101168521).