The code of the paper "Blind Video Super-Resolution based on Implicit Kernels".

Requirements

Python 3.9, PyTorch >= 1.9.1

Platforms: Ubuntu 22.04

Environment

conda create -n BVSR python=3.9 -y && conda activate BVSR

git clone --depth=1 https://github.com/QZ1-boy/BVSR && cd QZ1-boy/BVSR/

# given CUDA 11.1
python -m pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 torchaudio==0.9.1 -f https://download.pytorch.org/whl/torch_stable.html

python -m pip install tqdm lmdb pyyaml opencv-python scikit-image

Datasets

Training GT Datasets: REDS-GT.

Testing GT Datasets: REDS4-GT,Vid4-GT, UDM10-GT.

Testing Datasets on Gaussian Blur and Realistic Motion Blur: REDS4/Vid4/UDM10, Code [BVSR].

Put the downloaded training datasets and testing datasets into the ./dataset file path.

Pre-trained models: Moldes, Code [ckpt].

Train

CUDA_VISIBLE_DEVICES=1   python main_KCA.py  --train --Deg_option Gaussian_REDS --config_path exp_KCA_REDS_Gaussian.cfg 
CUDA_VISIBLE_DEVICES=1   python main_KCA.py  --train --Deg_option Realistic_REDS --config_path exp_KCA_REDS_Realistic.cfg 

Test

CUDA_VISIBLE_DEVICES=1   python main_KCA.py  --test_REDS4  --Deg_option Gaussian_REDS  --config_path exp_KCA_REDS_Gaussian.cfg
CUDA_VISIBLE_DEVICES=1   python main_KCA.py  --test_Vid4   --Deg_option Gaussian_REDS  --config_path exp_KCA_REDS_Gaussian.cfg 
CUDA_VISIBLE_DEVICES=1   python main_KCA.py  --test_UDM10  --Deg_option Gaussian_REDS  --config_path exp_KCA_REDS_Gaussian.cfg
CUDA_VISIBLE_DEVICES=1   python main_KCA.py  --test_REDS4  --Deg_option Realistic_REDS  --config_path exp_KCA_REDS_Realistic.cfg
CUDA_VISIBLE_DEVICES=1   python main_KCA.py  --test_Vid4   --Deg_option Realistic_REDS  --config_path exp_KCA_REDS_Realistic.cfg 
CUDA_VISIBLE_DEVICES=1   python main_KCA.py  --test_UDM10  --Deg_option Realistic_REDS  --config_path exp_KCA_REDS_Realistic.cfg

Citation

If this repository is helpful to your research, please cite our paper:

@article{zhu2025blind,
  title={Blind Video Super-Resolution based on Implicit Kernels},
  author={Zhu, Qiang and Jiang, Yuxuan and Zhu, Shuyuan and Zhang, Fan and Bull, David and Zeng, Bing},
  conference={International Conference on Computer Vision},
  year={2025}
}

Related Works

Our project was built on the video super-resolution method FMA-Net. We also release some blind video super-resolution models, e.g., DBVSR, BSVSR, Self-BVSR.