RefSR-18K

An LR-Conditioned Preference Dataset for Super-Resolution

 

Dataset Card

RefSR-18K is the first large-scale LR-conditioned preference dataset for super-resolution, introduced in "RefReward-SR: LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution". Each group provides pairwise human rankings based on LR–HR consistency and HR naturalness, rather than similarity to a ground-truth image.

The reward model trained on this dataset is released separately as casyssong/RefReward-SR.

Data Provenance

The images annotated in RefSR-18K originate from the LSDIR dataset. High-resolution source images from LSDIR were degraded to produce LR inputs, and multiple super-resolution methods were used to generate the HR candidates. Human annotators then ranked groups of four HR candidates according to their consistency with the LR input and their visual naturalness.

The CSV/JSONL files and per-annotator rankings in this repository contain the preference annotations collected by the RefReward-SR project.

Contents

Split	Groups	Size	Usage
train_images/	4499 (0001-4499)	12G	Used for reward model training
test_in_domain_images/	200	527M	In-domain benchmark
test_out_domain_images/	200	532M	Out-domain benchmark
train_extra_images/	3562 (4500-8061)	9.1G	Supplementary groups; not used in released training, provided for future research

Each image group contains LR.png (low-resolution input) and 4 SR outputs from different methods (HR_A.png, HR_B.png, HR_C.png, HR_D.png), with human-annotated pairwise rankings.

Annotations:

File	Description
train.csv	Annotations for train_images/ (4499 groups)
train_extra.csv	Annotations for train_extra_images/ (3562 groups)
train_sr.jsonl	Training file generated from train.csv + train_images/
test_in_domain.csv / test_out_domain.csv	Test set ground-truth rankings
raw-in-domain/ / raw-out-domain/	Per-annotator rankings, for inter-annotator agreement evaluation

Source Data License Information

This dataset card does not declare a new license for the RefSR-18K images or preference annotations. The applicable terms of the source image data are:

Source	Role in RefSR-18K	Upstream terms
LSDIR	Source of the images used to create the LR inputs and HR candidates that were ranked by annotators	Please notice that this dataset is made available for academic research purpose only. All collection and processing of data for LSDIR was performed by the academic co-authors. All the images are collected from the Internet, and the copyright belongs to the original owners. If any of the images belongs to you and you would like it removed, please kindly inform us, we will remove it from our dataset immediately.

The derived LR images and SR candidates originate from LSDIR images and remain subject to the applicable LSDIR terms and the rights of the original image owners. No separate license is assigned to the human preference annotations by this dataset card. Users are responsible for reviewing and complying with the upstream terms before using or redistributing any part of the dataset.

Citation

@article{song2026refreward,
  title={RefReward-SR: LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution},
  author={Song, Yushuai and Quan, Weize and Wang, Weining and Sun, Jiahui and Liu, Jing and Li, Meng and Yu, Pengbin and Chen, Zhentao and Shen, Wei and Yuan, Lunxi and Yan, Dong-ming},
  journal={arXiv preprint arXiv:2603.24198},
  year={2026}
}