HifiFace — Unofficial Pytorch Implementation

Image source: HifiFace: 3D Shape and Semantic Prior Guided High Fidelity Face Swapping (figure 1, pg. 1)

This repository is an unofficial implementation of the face swapping model proposed by Wang et. al in their paper HifiFace: 3D Shape and Semantic Prior Guided High Fidelity Face Swapping. This implementation makes use of the Pytorch Lighting library, a light-weight wrapper for PyTorch.

HifiFace Overview

The task of face swapping applies the face and the identity of the source person to the head of the target.

The HifiFace architecture can be broken up into three primary structures. The 3D shape-aware identity extractor, the semantic facial fusion module, and an encoder-decoder structure. A high-level overview of the architecture can be seen in the image below.

Image source: HifiFace: 3D Shape and Semantic Prior Guided High Fidelity Face Swapping (figure 2, pg. 3)

Changes from the original paper

Dataset

In the paper, the author used VGGFace2 and Asian-Celeb as the training dataset. Unfortunately, the Asian-Celeb dataset can only be accessed with a Baidu account, which we do not have. Thus, we only use VGGFace2 for our training dateset.

Model

The paper proposes two versions of HifiFace model based on the output image size: 256x256 and 512x512 (referred to as Ours-256 and Ours-512 in the paper). The 512x512 model uses an extra data preprocessing before training. In this open source project, we implement the 256x256 model. For the discriminator, the original paperuses the discriminator from StarGAN v2. Our implementation uses the multi-scale discriminator from SPADE.

Installation

Build Docker Image

git clone https://github.com/mindslab-ai/hififace 
cd hififace
git clone https://github.com/sicxu/Deep3DFaceRecon_pytorch && git clone https://github.com/NVlabs/nvdiffrast && git clone https://github.com/deepinsight/insightface.git
cp -r insightface/recognition/arcface_torch/ Deep3DFaceRecon_pytorch/models/
cp -r insightface/recognition/arcface_torch/ ./model/
rm -rf insightface
cp -rf 3DMM/* Deep3DFaceRecon_pytorch
mv Deep3DFaceRecon_pytorch model/
rm -rf 3DMM
docker build -t hififace:latent .
rm -rf nvdiffrast

This Dockerfile was inspired by @yuzhou164, this issue from Deep3DFaceRecon_pytorch.

Pre-Trained Model for Deep3DFace PyTorch

Follow the guideline in Prepare prerequisite models

Set up at ./mode/Deep3DFaceRecon_pytorch/

Pre-Trained Models for ArcFace

We used official Arcface per-trained pytorch implementation Download pre-trained checkpoint from onedrive (IResNet-100 trained on MS1MV3)

Download HifiFace Pre-Trained Model

google drive link trained on VGGFace2, 300K iterations

Training

Dataset & Preprocessing

Align & Crop

We aligned the face images with the landmark extracted by 3DDFA_V2. The code will be added.

Face Segmentation Map

After finishing aligning the face images, you need to get the face segmentation map for each face images. We used face segmentation model that PSFRGAN provides. You can use their code and pre-trained model.

Dataset Folder Structure

Each face image and the corresponding segmentation map should have the same name and the same relative path from the top-level directory.

face_image_dataset_folder
└───identity1
│   │   image1.png
│   │   image2.png
│   │   ...
│   
└───identity2
│   │   image1.png
│   │   image2.png
│   │   ...
│ 
|   ...

face_segmentation_mask_folder
└───identity1
│   │   image1.png
│   │   image2.png
│   │   ...
│   
└───identity2
│   │   image1.png
│   │   image2.png
│   │   ...
│ 
|   ...

Wandb

Wandb is a powerful tool to manage your model training. Please make a wandb account and a wandb project for training HifiFace with our training code.

Changing the Configuration

• config/model.yaml

  • dataset.train.params.image_root: directory path to the training dataset images
  • dataset.train.params.parsing_root: directory path to the training dataset parsing images
  • dataset.validation.params.image_root: directory path to the validation dataset images
  • dataset.validation.params.parsing_root: directory path to the validation dataset parsing images

• config/trainer.yaml

  • checkpoint.save_dir: directory where the checkpoints will be saved
  • wandb: fill out your wandb entity and project name

Run Docker Container

docker run -it --ipc host --gpus all -v /PATH_TO/hififace:/workspace -v /PATH_TO/DATASET/FOLDER:/DATA --name hififace hififace:latent

Run Training Code

python hififace_trainer.py --model_config config/model.yaml --train_config config/trainer.yaml -n hififace

Inference

Single Image

python hififace_inference --gpus 0 --model_config config/model.yaml --model_checkpoint_path hififace_opensouce_299999.ckpt --source_image_path asset/inference_sample/01_source.png --target_image_path asset/inference_sample/01_target.png --output_image_path ./01_result.png

All Posible Pairs of Images in Directory

python hififace_inference --gpus 0 --model_config config/model.yaml --model_checkpoint_path hififace_opensouce_299999.ckpt  --input_directory_path asset/inference_sample --output_image_path ./result.png

Interpolation

# interpolates both the identity and the 3D shape.
python hififace_inference --gpus 0 --model_config config/model.yaml --model_checkpoint_path hififace_opensouce_299999.ckpt --source_image_path asset/inference_sample/01_source.png --target_image_path asset/inference_sample/01_target.png --output_image_path ./01_result_all.gif  --interpolation_all 

# interpolates only the identity.
python hififace_inference --gpus 0 --model_config config/model.yaml --model_checkpoint_path hififace_opensouce_299999.ckpt --source_image_path asset/inference_sample/01_source.png --target_image_path asset/inference_sample/01_target.png --output_image_path ./01_result_identity.gif  --interpolation_identity

# interpolates only the 3D shape.
python hififace_inference --gpus 0 --model_config config/model.yaml --model_checkpoint_path hififace_opensouce_299999.ckpt --source_image_path asset/inference_sample/01_source.png --target_image_path asset/inference_sample/01_target.png --output_image_path ./01_result_3d.gif  --interpolation_3d

Our Results

The results from our pre-trained model.

GIF interpolaiton results from Obama to Trump to Biden back to Obama. The left image interpolates both the identity and the 3D shape. The middle image interpolates only the identity. The right image interpolates only the 3D shape.

To-Do List

• Pre-processing Code
• Colab Notebook

License

BSD 3-Clause License.

Implementation Author

Changho Choi @ MINDs Lab, Inc. ([email protected])

Matthew B. Webster @ MINDs Lab, Inc. ([email protected])

Citations

@article{DBLP:journals/corr/abs-2106-09965,
  author    = {Yuhan Wang and
               Xu Chen and
               Junwei Zhu and
               Wenqing Chu and
               Ying Tai and
               Chengjie Wang and
               Jilin Li and
               Yongjian Wu and
               Feiyue Huang and
               Rongrong Ji},
  title     = {HifiFace: 3D Shape and Semantic Prior Guided High Fidelity Face Swapping},
  journal   = {CoRR},
  volume    = {abs/2106.09965},
  year      = {2021}
}