VCISR: Blind Single Image Super-Resolution with Video Compression Synthetic Data (WACV 2024)

:star:If you like VCISR, please help star this repo. Thanks!:hugs:

:book:Table Of Contents

• Update
• Installation
• Train
• Inference
• Anime
• VC-RealLQ

Update

• 2023.12.08: The pre-trained weight is released.
• 2023.11.29: This repo is released.

Installation (Environment Preparation)

git clone git@github.com:Kiteretsu77/VCISR-official.git
cd VCISR

# Create conda env
conda create -n VCISR python=3.10
conda activate VCISR

# Install Pytorch we use torch.compile in our repository by default
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

# Install FFMPEG (the following is for linux system, the rest can see https://ffmpeg.org/download.html)
sudo apt install ffmpeg

Train

1. Download Datasets (DIV2K) and crop them by the script below (following our paper):

   shell bash scripts/download_datasets.sh

2. Train: Please check opt.py to setup parameters you want\ Step1 (Net L1 loss training): Run shell python train_code/train.py The model weights will be inside the folder 'saved_models'

   Step2 (GAN Adversarial Training): 1. Change opt['architecture'] in opt.py as "GRLGAN". 2. Rename weights in 'saved_models' (either closest or the best, we use closest weight) to grlgan_pretrained.pth 3. Run shell python train_code/train.py --use_pretrained

Inference:

1. Download the weight from https://drive.google.com/file/d/1Mbrw1ji_qcOteuSOkZqVgSEda_PQ40tA/view?usp=drive_link and put them in "saved_models" folder
2. Setup the configuration of test_code/inference.py after line 215.
3. Then, Execute shell python test_code/inference.py

Anime:

We also extend our methods on the Anime Restoration and Super-Resolution task with public and private Anime datasets. \ You can also find a pre-built highly accelerated Anime SR inference repository from: \ https://github.com/Kiteretsu77/Anime_SR_Restoration (A regular inference tool) or \ https://github.com/Kiteretsu77/FAST_Anime_VSR (An accelerated processing repository).\ These two repositories are RRDB-based network training (instead of GRL). \ In the future, we will release more details related to the Anime extension and versatile pre-trained weights for different models, such as GRL.

VC-RealLQ:

The small image inference dataset will be released soon. If you need it earlier, you can contact hikaridawn412316@gmail.com.

Citation

Please cite us if our work is useful for your research.

@article{wang2023vcisr,
  title={VCISR: Blind Single Image Super-Resolution with Video Compression Synthetic Data},
  author={Wang, Boyang and Liu, Bowen and Liu, Shiyu and Yang, Fengyu},
  journal={arXiv preprint arXiv:2311.00996},
  year={2023}
}

License

This project is released under the GPL 3.0 license.

Contact

If you have any questions, please feel free to contact me at hikaridawn412316@gmail.com or boyangwa@umich.edu.