IA-FaceS: A Python repository from CMACH508

IA-FaceS — Official PyTorch Implementation

This repository contains the supplementary material and the official PyTorch implementation of the paper:

IA-FaceS: A Bidirectional Method for Semantic Face Editing

Abstract: Semantic face editing has achieved substantial progress in recent years. However, existing face editing methods, which often encode the entire image into a single code, still have difficulty in enabling flexible editing while keeping high-fidelity reconstruction. The one-code scheme also brings entangled face manipulations and limited flexibility in editing face components. In this paper, we present IA-FaceS, a bidirectional method for disentangled face attribute manipulation as well as flexible, controllable component editing. We propose to embed images onto two branches: one branch computes high- dimensional component-invariant content embedding for capturing face details, and the other provides low-dimensional component-specific embeddings for component manipulations. The two-branch scheme naturally enables high-quality facial component-level editing while keeping faithful reconstruction with details. Moreover, we devise a component adaptive modulation (CAM) module, which integrates component- specific guidance into the decoder and successfully disentangles highly-correlated face components. The single-eye editing is developed for the first time without editing face masks or sketches. According to the experimental results, IA-FaceS establishes a good balance between maintaining image details and performing flexible face manipulation. Both quantitative and qualitative results indicate that the proposed method outperforms the existing methods in reconstruction, face attribute manipulation, and component transfer.

Demo video and supplementary file

Supplementary materials related to our paper are available at the following links:

Path	Description
supplementary_material.pdf	Supplementary file for IA-FaceS.
IA-FaceS_demo.mp4	The video demo of IA-FaceS.

Installation

Install the dependencies:

conda create -n iafaces python=3.7
conda activate iafaces
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
pip install -r requirements.txt

For docker users:

docker pull huangwenjingcs/ubuntu18-conda-cuda11-pytorch1.7

Datasets and pre-trained networks

To obtain the CelebA-HQ dataset, please refer to the Progressive GAN repository. The official way of generating CelebA-HQ can be challenging. You can get the pre-generated dataset from CelebA-HQ-dataset. Unzip the file and put the images to "data/CelebA-HQ-img/".

To obtain the Flickr-Faces-HQ Dataset (FFHQ), please refer to ffhq-dataset. Download the dataset and put the images to "data/ffhq-images1024/".

Pre-trained models can be found via the following links:

Path	Description
checkpoint	Main folder.
├ iafaces-celebahq-256.pth	IA-FaceS trained with CelebA-HQ dataset at 256×256.
├ iafaces_cam-celebahq-256.pth	IA-FaceS-CAM trained with CelebA-HQ dataset at 256×256.
├ iafaces-ffhq-1024.pth	IA-FaceS trained with FFHQ dataset at 1024×1024.
├ iafaces_cam-ffhq-1024.pth	IA-FaceS-CAM trained with FFHQ dataset at 1024×1024.

Download the pre-trained networks and put them to "iafaces-evacheckpoints/".

Train networks

Once the datasets are set up, you can train the networks as follows:

Edit configs/<EXP_ID>.json to specify the dataset, model and training configurations.
Run the training script with python train.py -c configs/<EXP_ID>.json . For example,

 # train IA-FaceS with CelebA-HQ (256px) dataset, with a batch size of 16
python train.py -c configs/iafaces-celebahq-256.json --bz 16

The code will use all GPUS by default, please specify the devices you want to use by:

 # train IA-FaceS in parallel, with a batch size of 16 (paper setting on celebahq)
CUDA_VISIBLE_DEVICES=0,1 python train.py -c configs/iafaces-celebahq-256.json --bz 8

The checkpoints are written to a newly created directory saved/models/<EXP_ID>

Edit images

If you want to edit images using pre-trained models, go into the "iafaces-eval/" folder.

For attribute manipulation,

To compute the attribute direction, run:

# collect latent codes
python collect_latent_codes.py --data_path <DATA_LIST> --img_dir <IAMGE_DIR> --resume <CHECKPOINT>
python train_boundary.py --index <COMPONENT INDEX> --pose_attr <POSITIVE ATTRIBUTE> --neg_attr <NEGTIVE ATTRIBUTE> --expr <EXP ID>

To edit face attribute, run:

python attr_manipulation.py --attr <ATTRIBUTE> --data_path <DATA_LIST> --img_dir <IAMGE_DIR> --resume <CHECKPOINT>

For face component transfer, run:

python component_transfer.py --component <COMPONENT> --target <TARGET_LIST> --reference <REFERENCE_LIST> --img_dir <IAMGE_DIR> --resume <CHECKPOINT>

For image reconstruction, run:

python reconstruction.py --data_path <DATA_LIST> --img_dir <IAMGE_DIR> --resume <CHECKPOINT>

The results are saved to output/.

If you want to see details, please follow iafaces-eval/README.md.

Metrics

Reconstruction: MSE, LPIPS, PSNR, SSIM, FID
Component transfer: MSE$_{\text{irr}}$, IFG, FID
Attribute manipulation: MSE$_{\text{irr}}$, IFG, Arc-dis

If you want to see details, please follow metrics/README.md.

Citation

If you find this work useful for your research, please cite our paper:

@article{huang2022ia,
  title={IA-FaceS: A bidirectional method for semantic face editing},
  author={Huang, Wenjing and Tu, Shikui and Xu, Lei},
  journal={Neural Networks},
  year={2022},
  publisher={Elsevier}
}

Acknowledgement

This repository used some codes in pytorch-template and stylegan2-pytorch.

CMACH508/IA-FaceS