RefReward-SR

LR-Conditioned Reward Modeling for Preference-Aligned Super-Resolution

 

Model Card

This repository contains the released RefReward-SR checkpoint only. The SR-GRPO-LoRA weights are released in a separate repository.

RefReward-SR is an LR-conditioned reward model for super-resolution. Given a low-resolution (LR) reference image and a super-resolved (HR) image, it assesses semantic consistency and perceptual plausibility using the LR image as a semantic anchor. The model supports direct scoring (fast) and reasoning-aware scoring (think).

Checkpoint

Item	Description	Size
RefReward-SR (checkpoint-1200)	Qwen3-VL-8B LR-conditioned reward model fine-tuned with GRPO	17G

Base Weight and Training Dataset

Type	Resource	Role
Base weight	Qwen3-VL-8B-Instruct	Initialization for GRPO fine-tuning
Training dataset	RefSR-18K	LR-conditioned human preference data used for reward-model training

Upstream License Information

This model card does not declare a new license for the released checkpoint. The relevant upstream terms are:

Resource	Upstream license
Qwen3-VL-8B-Instruct	Apache License 2.0
RefSR-18K	See the RefSR-18K dataset card

Users are responsible for reviewing and complying with the applicable upstream terms.

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}
}