FontGenerator

Alphabet font generator using GAN, inspired from zi2zi, pix2pix and other similar GAN projects. (Experimented in Windows11 / RTX2070super / RAM 16GB)

Project Objective

The objective of out project is to make a GAN model that learns font styles from other pre-existing fonts. Making a new font style is a very difficult task but generative adversarial network (GAN) can reduce the time cost. There are several prior researches proving that GAN is a good style transfer model. We trained the GAN model to generate a font image, where the generator creates a fake image and the discriminator distinguishes whether it is generated. With pre-training and fine-tuning the model, we were able to train the model to learn the font styles. Through this, it can be assured that GAN has a potential to simplify the font style production work by creating the entire font data with only a small number of font data.

Implementation

Reproduction

Set conda environment

$ cd ./FontGenerator
$ conda env create --file conda_requirements.yaml
$ conda activate FontProject

Make Directory

$ mkdir ./results
$ mkdir ./results/ckpt
$ mkdir ./results/ckpt/pre_train
$ mkdir ./results/ckpt/finetune
$ mkdir ./results/fake-image/pre_train
$ mkdir ./results/fake-image/pre_train/valid
$ mkdir ./results/fake-image/finetune
$ mkdir ./results/fake-image/finetune/valid
$ mkdir ./train/dataset

1. Install Font files and convert to images

https://drive.google.com/file/d/1iRYDXJbH_x4Kabr52LudkvE8JnspLaN6/view?usp=sharing
Extract font files in Util/Font
$ python Util/FontImageGerator.py

2. Install Pretrained model files

https://drive.google.com/file/d/1nu4fKGHy5HVb_s_k425R8LquvBHgXxVJ/view?usp=share_link
Extract model files in ./results/ckpt/pre_train

3. Pretrain

If trying to reproduce pretrain model If don't trying to train and just use trained model parameters pass.

$ python ./train/train.py

4. Pretrain Validation(Inference)

$ python ./train/valid.py

validation results are in results/pre_train/valid-.png Target results are results/pre_train/valid-true-.png

5. Finetuning

$ python ./train/finetune.py

6. Result Images

Check images in ./results/fake-image/finetune

7. Evaluation Table

$ python ./main.py

Training Results

	Train	Validation	Finetuning
L1 loss (x100)	0.3913	1.1322	0.0674

Training(80%) and Validation(20%)

Training Loss

Finetunning Results - Italic Calibri

References

• zi2zi
• zi2zi-blog
• GAN-handwriting-styler
• GAN-handwriting-styler-blog
• GAN-goodfellow