ExFaceGAN: A Python repository from fdbtrs

ExFaceGAN: Exploring Identity Directions in GAN’s Learned Latent Space for Synthetic Identity Generation

Accepted at IJCB 2023 Arxiv

Requirements for ExFaceGAN

It is recommended to create a virtual environment with Python 3.6 and requirements.txt

Download pretrained StyleGAN generator models from this link while following their license and place stylegan3-r-ffhq-1024x1024.pkl and stylegan2-ffhq-1024x1024.pkl in generators/pretrain/.

Download the pretrained GANControl model from this link while following their license and extract it in generators/GANControl_resources/gan_models/.

Pipeline for Disentangling Identity Information and Generate Images

Generate unconditional images
Create SVM training data
Train SVMs
Generate facial images
Train FR model

1. Generate Unconditional Images

CUDA_VISIBLE_DEVICES=0 python generate_SG_imgs.py --num_imgs 11000 --modelname "stylegan3_ffhq" --save_path "path/to/save/images_and_latents"

Align images via:

CUDA_VISIBLE_DEVICES=0 python MTCNN_alignment.py --in_folder "path/to/images_and_latents/images" --out_folder "path/to/images_and_latents/images_aligned"

2. Create SVM Training Data

python create_boundary_data.py --datadir "path/to/previous/images_and_latents" --save_path "path/to/save/boundary/data" --fr_path "path/tp/pretrained/FR/model"

3. Train SVMs

python train_boundaries.py --datadir "path/to/previous/boundary/data" --save_path "path/to/save/boundaries"

4. Generate Facial Images

CUDA_VISIBLE_DEVICES=0 python create_dataset.py --latent_dir "path/to/previous/images_and_latents/w_latents" --boundary_dir "path/to/previous/boundaries" --modelname "stylegan3_ffhq" --output_path "path/to/save/new/dataset"

or set these parameters as default and call create_dataset.sh (change num_classes and offset) to generate images on 4 GPUs in parallel.

5. Train FR Model

Change config_baseline.py and run.sh to your preferences and execute:

run.sh

Additional Scripts

visualization/analyse_dataset.py plots genuine-impostor-distributions and saves genuine and impostor scores for given datasets.

Given genuine, impostor scores, visualization/dataset_EER.sh calculates multiple metrics including EER, FMR, ...

visualization/impact_of_max_off.py allows editing one latent code and plots cosine similarity between edited images and the reference.

Citation

If you use any of the code provided in this repository or the models provided, please cite the following paper:

@InProceedings{Boutros_2023_IJCB,
    author    = {Boutros, Fadi and Klemt, Marcel and Fang, Meiling and Arjan Kuijper and Damer, Naser},
    title     = {ExFaceGAN: Exploring Identity Directions in GAN’s Learned Latent Space for
Synthetic Identity Generation},
    booktitle = {{IEEE} International Joint Conference on Biometrics, {IJCB} 2023},
    month     = {September},
    year      = {2023}
}

fdbtrs/ExFaceGAN