JointFontGAN in PyTorch

This is the implementation of our ACM International Conference on Multimedia 2020 paper "JointFontGAN: Joint Geometry-Content GAN for Font Generation via Few-Shot Learning". The code was written by Yankun Xi. More details are given in the following.

Prerequisites:

• Linux or macOS
• Python 3.6 or later (latest built on Python 3.8)
• Pytorch 1.2 or later (latest built on Pytorch 1.6)
• CPU or NVIDIA GPU + CUDA CuDNN

Getting Started

Installation

pip install visdom
pip install dominate
pip install scikit-image

• Clone this repo:

mkdir JointFontGAN
cd JointFontGAN
git clone https://github.com/yankunxi/JointFontGAN
mkdir dataset

• Download dataset:

Download the following two font datasets into dataset folder and unzip. Each of the datasets consists of training and test images.

Capitals64 dataset: https://drive.google.com/file/d/1qrxhhgG2vwUhhq-shbHxzt3b1ahkoNt_/view?usp=sharing

SandunLK10k64 dataset: https://drive.google.com/file/d/1VgzxiBrYYUdB0eyNKVb137W0jY43YCeM/view?usp=sharing

• Enter this repo:

cd JointFontGAN
mkdir checkpoints

• (Optional) Download pre-trained model

Download the following models into checkpoints folder and unzip.

Capitals54 dataset: https://drive.google.com/file/d/1C3JvbjdRecqVc3UmWxR1mLP_i0hwmDxp/view?usp=sharing

SandunLK10k64 dataset: https://drive.google.com/file/d/1T140Uig4CfL8W6vsh0TElkguBAJrf_Rp/view?usp=sharing

JointFontGAN train/test

• To train the model, please run the following scripts for the two datasets:

. ./scripts/EskGAN/XItrain_EskGAN.sh Capitals64

. ./scripts/EskGAN/XItrain_EskGAN2_dspostG=1.sh SandunLK10k64

Or you can skip the training phase and test on our pretrained models.

• To test the model:

. ./scripts/EskGAN/XItest_EskGAN.sh Capitals64 test

. ./scripts/EskGAN/XItest_EskGAN2_dspostG=1.sh SandunLK10k64 test

• We also provide our generated test font results:

Capitals54 dataset: https://drive.google.com/file/d/1gjqnjhdes2rsTr6bX3rpGsBaWw_sIEyn/view?usp=sharing

SandunLK10k64 dataset: https://drive.google.com/file/d/118hPUy2jRHn7wRZTYcDhfuOJJsvnbdLF/view?usp=sharing

• GPU difference:

Based on different GPU RAM, two parameters might need to be modified in the training scripts. Generally, with less RAM, one would like to use smaller BATCHSIZE, but keep the product BATCHSIZE * BATCHSPLIT unchanged.

Citation

If you use this code for your research, please cite:

@inproceedings{xi2020jointfontgan,
  title={JointFontGAN: Joint Geometry-Content GAN for Font Generation via Few-Shot Learning},
  author={Xi, Yankun and Yan, Guoli and Hua, Jing and Zhong, Zichun},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={4309--4317},
  year={2020}
}

Acknowledgements

Code is inspired by MC-GAN. Datasets are collected from MC-GAN and Sandun.LK