Type design and artificial intelligence

As I wrote on this blog post, I started to experiment with typography using artificial intelligence. the first idea was to encode font’s path into text (using svg bezier encoding) and parse these paths as data for a text generative models such as gpt-2. It shows its limit by giving strange paths or draw text too similar to the database.

So I went with GAN (generative adversial network) which, from a dataset of bitmap image can generate new image. So the font need to be converted to an image each letter in a specific delimited size image. I choose 256x256px and 8 bit RGB (BW format and RGBA not supported by lots of script). And I specified custom data into a json format to get it as a conditionnal GAN ( where I can precise latent space information for my data … ), it sounds a bit complex but it allows me to make a generative model where I can precise: draw me a “A” or an “@” !

I made only “Hamburgerfon” type — since I had issue training the model with all the font. Here are the result after few days of training :