G F P G A Nv1

Introduction

GFPGAN (Generative Facial Prior GAN) is a blind face restoration algorithm that utilizes generative facial priors to enhance the quality and realism of face images, particularly when dealing with low-quality inputs. It achieves high-quality restoration by incorporating rich priors through a pretrained face GAN.

Architecture

GFPGAN leverages a Generative Facial Prior embedded in a pretrained GAN model. This is integrated into the restoration process using channel-split spatial feature transform layers. These layers allow a balance between realness and fidelity, enabling the restoration of facial details and color enhancement in a single forward pass.

Training

The training process for GFPGAN involves several steps:

1. Dataset Preparation: Use the FFHQ dataset.
2. Pre-trained Model Downloads: Obtain the StyleGAN2 model, component locations of FFHQ, and a simple ArcFace model.
3. Configuration: Modify the train_gfpgan_v1.yml file as needed.
4. Execution: Run the training using distributed PyTorch with the given command.

Guide: Running Locally

To run GFPGAN locally, follow these steps:

1. Clone the Repository:

   git clone https://github.com/xinntao/GFPGAN.git
   cd GFPGAN
   

2. Install Dependencies:

   • Install BasicSR and facexlib:

     BASICSR_EXT=True pip install basicsr
     pip install facexlib
     

   • Install other requirements:

     pip install -r requirements.txt
     

3. Download Pre-trained Models:

   wget https://github.com/TencentARC/GFPGAN/releases/download/v0.1.0/GFPGANv1.pth -P experiments/pretrained_models
   

4. Run Inference:

   python inference_gfpgan_full.py --model_path experiments/pretrained_models/GFPGANv1.pth --test_path inputs/whole_imgs
   

For optimal performance, use cloud GPUs like those offered by Google Colab or AWS.

License

GFPGAN is released under the Apache License Version 2.0.