SR-GRPO-LoRA

Preference-Aligned Super-Resolution

Model Card

This repository contains the released SR-GRPO-LoRA weights.

SR-GRPO-LoRA is a LoRA adapter for preference-aligned super-resolution, obtained by fine-tuning a FLUX.1-dev-based ControlNet SR model with GRPO.

For inference, load the released SR-GRPO-LoRA weights into the super-resolution pipeline built with FLUX.1-dev and the C-FLUX ControlNet checkpoint distributed by DP2O-SR.

Checkpoint

Item	Description	Size
SR-GRPO-LoRA	LoRA adapter for FLUX ControlNet SR, trained with GRPO	72M

Upstream Weights

Resource	Role
FLUX.1-dev	Base diffusion model and VAE
C-FLUX ControlNet checkpoint from DP2O-SR	Frozen ControlNet weights for super-resolution conditioning

Training Datasets

The LoRA was trained on degraded LR/HR pairs prepared from:

Dataset	Usage
LSDIR	Image-restoration training images
FFHQ	The first 10,000 images were used for training

The data were prepared with the SeeSR degradation pipeline to produce ground-truth images, degraded LR images, and tag prompts.

Upstream License Information

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

Weights

Resource	Upstream license
FLUX.1-dev	FLUX.1 [dev] Non-Commercial License
C-FLUX ControlNet checkpoint from DP2O-SR	Apache License 2.0

Datasets

Dataset	Upstream terms
LSDIR	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.
FFHQ	See the FFHQ license information

Users are responsible for reviewing and complying with all 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}
}