VQFR (ECCV 2022 Oral)

This paper aims at investigating the potential and limitation of Vector-Quantized (VQ) dictionary for blind face restoration.
We propose a new framework VQFR – incoporating the Vector-Quantized Dictionary and the Parallel Decoder. Compare with previous arts, VQFR produces more realistic facial details and keep the comparable fidelity.

VQFR: Blind Face Restoration with Vector-Quantized Dictionary and Parallel Decoder

[Paper] [Project Page] [Video] [B站] [Poster] [Slides]
Yuchao Gu, Xintao Wang, Liangbin Xie, Chao Dong, Gen Li, Ying Shan, Ming-Ming Cheng
Nankai University; Tencent ARC Lab; Tencent Online Video; Shanghai AI Laboratory;
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences

🔧 Dependencies and Installation

Python >= 3.7 (Recommend to use Anaconda or Miniconda)
PyTorch >= 1.7
Option: NVIDIA GPU + CUDA
Option: Linux

Installation

Clone repo

git clone https://github.com/TencentARC/VQFR.git
cd VQFR

Install dependent packages

# Build VQFR with extension
pip install -r requirements.txt
VQFR_EXT=True python setup.py develop

# Following packages are required to run demo.py

# Install basicsr - https://github.com/xinntao/BasicSR
pip install basicsr

# Install facexlib - https://github.com/xinntao/facexlib
# We use face detection and face restoration helper in the facexlib package
pip install facexlib

# If you want to enhance the background (non-face) regions with Real-ESRGAN,
# you also need to install the realesrgan package
pip install realesrgan

⚡ Quick Inference

Download pre-trained VQFR model [Google Drive|腾讯微云].

Inference

# for real-world image
python demo.py -i inputs/whole_imgs -o results -v 1.0 -s 2

# for cropped face
python demo.py -i inputs/cropped_faces/ -o results -v 1.0 -s 1 --aligned

Usage: python demo.py -i inputs/whole_imgs -o results -v 1.0 -s 2 [options]...

  -h                   show this help
  -i input             Input image or folder. Default: inputs/whole_imgs
  -o output            Output folder. Default: results
  -v version           VQFR model version. Option: 1.0. Default: 1.0
  -s upscale           The final upsampling scale of the image. Default: 2
  -bg_upsampler        background upsampler. Default: realesrgan
  -bg_tile             Tile size for background sampler, 0 for no tile during testing. Default: 400
  -suffix              Suffix of the restored faces
  -only_center_face    Only restore the center face
  -aligned             Input are aligned faces
  -ext                 Image extension. Options: auto | jpg | png, auto means using the same extension as inputs. Default: auto

💻 Training

We provide the training codes for VQFR (used in our paper).

Dataset preparation: FFHQ
Download lpips weights[Google Drive|腾讯微云] into experiments/pretrained_models/

Codebook Training

Pre-train VQ codebook on FFHQ datasets.

 python -m torch.distributed.launch --nproc_per_node=8 --master_port=2022 vqfr/train.py -opt options/train/VQGAN/train_vqgan_v1_B16_800K.yml --launcher pytorch

Or download our pretrained VQ codebook [Google Drive|腾讯微云] and put them in the experiments/pretrained_models folder.

Restoration Training

Modify the configuration file options/train/VQFR/train_vqfr_v1_B16_200K.yml accordingly.
Training

python -m torch.distributed.launch --nproc_per_node=8 --master_port=2022 vqfr/train.py -opt options/train/VQFR/train_vqfr_v1_B16_200K.yml --launcher pytorch

📏 Evaluation

We evaluate VQFR on one synthetic dataset CelebA-Test, and three real-world datasets LFW-Test, CelebChild and Webphoto-Test. For reproduce our evaluation results, you need to perform the following steps:

Download testing datasets (or VQFR results) by the following links:

Name	Datasets	Short Description	Download	VQFR Results
Testing Datasets	CelebA-Test(LQ/HQ)	_{3000 (LQ, HQ) synthetic images for testing}	Google Drive / 腾讯微云	Google Drive / 腾讯微云
	LFW-Test(LQ)	_{1711 real-world images for testing}
	CelebChild(LQ)	_{180 real-world images for testing}
	Webphoto-Test(LQ)	_{469 real-world images for testing}

Install related package and download pretrained models for different metrics:

    # LPIPS
    pip install lpips

    # Deg.
    cd metric_paper/
    git clone https://github.com/ronghuaiyang/arcface-pytorch.git
    mv arcface-pytorch/ arcface/
    rm arcface/config/__init__.py arcface/models/__init__.py

    # put pretrained models of different metrics to "experiments/pretrained_models/metric_weights/"

Metrics	Pretrained Weights	Download
FID	inception_FFHQ_512.pth	Google Drive / 腾讯微云
Deg	resnet18_110.pth
LMD	alignment_WFLW_4HG.pth

Generate restoration results:

Specify the dataset_lq/dataset_gt to the testing dataset root in test_vqfr_v1.yml.
Then run the following command:

    python vqfr/test.py -opt options/test/VQFR/test_vqfr_v1.yml

Run evaluation:

    # LPIPS|PSNR/SSIM|LMD|Deg.
    python metric_paper/[calculate_lpips.py|calculate_psnr_ssim.py|calculate_landmark_distance.py|calculate_cos_dist.py]
    -restored_folder folder_to_results -gt_folder folder_to_gt

    # FID|NIQE
    python metric_paper/[calculate_fid_folder.py|calculate_niqe.py] -restored_folder folder_to_results

📜 License

VQFR is released under Apache License Version 2.0.

👀 Acknowledgement

Thanks to the following open-source projects:

Taming-transformers

GFPGAN

DistSup

📋 Citation

@inproceedings{gu2022vqfr,
  title={VQFR: Blind Face Restoration with Vector-Quantized Dictionary and Parallel Decoder},
  author={Gu, Yuchao and Wang, Xintao and Xie, Liangbin and Dong, Chao and Li, Gen and Shan, Ying and Cheng, Ming-Ming},
  year={2022},
  booktitle={ECCV}
}

📧 Contact

If you have any question, please email yuchaogu9710@gmail.com.

ofirkris / VQFR