📄 Paper Link

Ce Wang¹, Wanjie Sun¹

¹ School of Remote Sensing and Information Engineering, Wuhan University

• Visual Results
• Installation
• Pretrained Models
• Dataset
• Train
• Inference
• Citation
• Acknowledgements
• Contact

# Clone this repository
git clone https://github.com/wwangcece/CRefDiff.git

# Create a conda environment with Python >= 3.9
conda create -n CRefDiff python=3.9
conda activate CRefDiff

# Install required packages
pip install -r requirements.txt

Download the pretrained models from the link below and place them in the checkpoints/ directory:

Download from HuggingFace

1. Refer to the Real-RefRSSRD for downloading.
2. Use the script dataset/prepare_lr.py to upscale the LR images to match the size of HR.

Firstly load pretrained SD parameters:

python scripts/init_weight_refsr.py \
--cldm_config configs/model/refsr_dino.yaml \
--sd_weight checkpoints/v2-1_512-ema-pruned.ckpt \
--output checkpoints/init_weight/init_weight-refsr.pt

Secondly please modify the training configuration files at configs/train_refsr.yaml. Finally you can start training:

python train.py \
--config configs/train_refsr.yaml

1. Modify the validation dataset configuration in configs/dataset/reference_sr_test.yaml and update the pretrained model path in inference_refsr_batch.py.
2. Run the inference script:

python inference_refsr_batch.sh --ckpt path/to/pretrained/model --output path/tp/out/dir --global_ref_scale 1 --device cuda:0 

If you find this work helpful, please consider citing:

@misc{wang2025controllablereferencebasedrealworldremote,
      title={Controllable Reference-Based Real-World Remote Sensing Image Super-Resolution with Generative Diffusion Priors}, 
      author={Ce Wang and Wanjie Sun},
      year={2025},
      eprint={2506.23801},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.23801}, 
}

This project is based on DiffBIR. We thank the authors for their excellent work.

If you have any questions, feel free to reach out to: Ce Wang — cewang@whu.edu.cn