Reimplementation of Progressive Semantic-Aware Style Transformation for Blind Face Restoration

Full explanation of this project can be found here.

Prerequisite

1. Python 3.7
2. CUDA 10.1
3. torch==1.5.1
4. torchvision==0.6.1
5. tensorflow>=1.15.4
6. opencv-python
7. dlib
8. scipy==1.4.1
9. tqdm
10. imgaug
11. torchmetrics
12. pytorch-fid

Training

The script to train the model using FFHQ dataset is available at train.py. The config for running this script can be found at config/base.json and config/psfrgan/train.json.

Downsample Test Dataset

The script to generate low-resolution (LR) face images and ground truth high-resolution (HR) face images from 1024x1024 CelebAHQ dataset for testing purpose is available at downsample_psfrgan.py. The config for running this script can be found at config/downsample.json.

Testing

The script to generate the predicted super-resolution (SR) face images is available at test_psfrgan.py. The config for running this script can be found at config/base.json and config/psfrgan/test.json.

Evaluation

The script to evaluate the predicted SR face images against ground truth HR face images is available at evaluate_psfrgan.py. The config for running this script can be found at config/evaluate.json. The script will produce PSNR, LPIPS, and SSIM score. To retrieve FID score, simply run python -m pytorch_fid <path/to/sr-folder> <path/to/hr-folder>.

Test your Own Image

The script to prepare your own test images from real LR face images is available at preprocess_psfrgan.py. This script will crop and align the real faces so that it is ready to be used to evaluate the model. The config for running this script is available at config/preprocess.json.