Official PyTorch implementation of GANiry.
GANiry: Bald-to-Hairy Translation Using CycleGAN,
Fidan Samet, Oguz Bakir.
(arXiv pre-print)
This work presents our computer vision course project called bald men-to-hairy men translation using CycleGAN. On top of CycleGAN architecture, we utilize perceptual loss in order to achieve more realistic results. We also integrate conditional constrains to obtain different stylized and colored hairs on bald men. We conducted extensive experiments and present qualitative results in this work.
-
Create new conda environment
conda create --name ganiry -
Activate the environment
conda activate ganiry -
Install the requirements
pip install -r requirements.txt -
Download CelebA dataset and prepare sub-dataset
python build_copy.py --dataroot ./datasets/bald2hairy --celeba_path ./datasets/celeba/data
Pre-trained models are also available.
Number of classes indicates the different hair classes in the dataset.
python train.py --dataroot ./datasets/bald2hairy --name bald2hairy --no_dropout --netG resnet_6blocks --load_size 143 --crop_size 128 --input_nc 4 --class_num 4 --percept_loss True --cycle_loss False
One hot vector is the binary encoding of hair classes.
python test.py --dataroot ./datasets/bald2hairy --name bald2hairy --no_dropout --netG resnet_6blocks --load_size 143 --crop_size 128 --input_nc 4 --class_num 4 --percept_loss True --cycle_loss False --phase test --one_hot_vector 1 0 1 0
GANiry is released under GNU General Public License. We developed GANiry on top of CycleGAN. Please refer to License of CycleGAN for more details.
If you find GANiry useful for your research, please cite our paper as follows.
F. Samet, O. Bakir, "GANiry: Bald-to-Hairy Translation Using CycleGAN", arXiv, 2021.
BibTeX entry:
@misc{samet2021ganiry,
title={GANiry: Bald-to-Hairy Translation Using CycleGAN},
author={Fidan Samet and Oguz Bakir},
year={2021},
eprint={2109.13126},
archivePrefix={arXiv},
primaryClass={cs.CV}
}