Font creation using 2d GAN

• kaggle

• wandb

• Colab
• nbviewer

Goal

Main goal of this project is to generate new fonts using GANs. The idea is to choose several images from the google fonts as a dataset, and train a GAN on it.

Synthetic data generation

We use 746 google fonts to generate synthetic data. Each font consists of 28x28 images of all 26 alphabets.

Model

We train a GAN with 6000 epochs, described in the notebook. We use the architecture from this repository, and modify it to suit our needs.

Result

Generated fonts at the end of 6000th epoch are shown in the notebook.

Image result

Losses

Interesting runs

An interesting run where G loss went to 0: link

This indicates mode collapse, and would be interesting to investigate further.