Skip to content

Re-implementation for ligoudaner377/font_translator_gan for our team project (CS492I, KAIST 2022F)

Notifications You must be signed in to change notification settings

deepkyu/multilingual-font-style-transfer

Repository files navigation

Font Style Transfer across multiple languages

This is an unofficial implementation for the paper Few-shot Font Style Transfer between Different Languages. The original implementation from the author can be found here.

News

🤗 Hugging Face Space demo is out!

deepkyu/multilingual-font-style-transfer

image

2023.09.29 I newly update the code and train the model to support Korean, Japanese, and Chinese characters. The training code has not yet updated in this repo since I have no time to integrate with this code. But I made a Hugging Face Space to show you whether the current "Multilingual Font Style Transfer" makes sense or not. I'll update whenever possible til we can get more established font style transfer model which can support the next multilingual society!

Setup

Requirements

See requirements.txt

Checkpoints

Download pretrained checkpoints for the evaluation to pretrained directory.

Data

For FTransGAN original data, you can download at the official repository.

For our multilingual dataset, you can download at our google drive link.

Please locate in data/ftgan-fonts and data/google-fonts for each.

Cheatsheet

Train

python trainer.py

Evaluation

For the result of FTrasnGAN and its data,

bash evaluate-ftgan.sh

The default option is with test_unknown_content. You can change it in config/datasets/ftgan.yaml by switching the value from True to False. (If False, it evaluates with test_unknown_style.)

For the result of Google Fonts data,

bash evaluate-google-font.sh

Updates

2022.12.04

This code is a part of my final term project in Special Topics in Computer Science - Introduction to Deep Learning (CS492I) (2022 Fall, KAIST).
The goal of my proposal is to generate novel fonts for a certain language, whose style is transferred from fonts for other languages.
In this project, I used google/font repository to define Google Fonts dataset, which contains font images from 10 different languages.

However, I failed to introduce the font generation models for unseen languages. Instead, I'll share some results in training dataset to explain the core concept of this project.

Hope that I could finish this project successfully in some days :)

Main concept

evaluation-result overfitted-result

To be continued...

  • finish the military service in South Korea
  • finish the undergraduate course and (might) go to graduate school...

About

Re-implementation for ligoudaner377/font_translator_gan for our team project (CS492I, KAIST 2022F)

Topics

Resources

Stars

Watchers

Forks