Logograms GAN Save

Generating logogams from the movie "Arrival" using generative adversarial networks in Keras

Project README

Recently I watched the movie Arrival(about an alien language) and did some reading about the language used in the film.

The symbols of the alien language(called logograms) are very elegant:

The creators uploaded some of the symbols to github for public use. I used these 40 logograms and trained a Generative Adversarial Network(GAN) on them. I coded a GAN in Keras (with a lot of help from other projects on github) and modified it a bit to work for logograms. In the end I was able to generate 128x128 pixel images:

These images are certainly not as clean as the originals but they are unmistakably similar. I wrote a simple script to clean up these images by removing stray marks.

Unfortunately many of these images look like copies of the original logograms. I believe this is because of the small number of samples. GAN's are also notoriously hard to train. Hopefully the producers will release more of the language in the future. Some future work I hope to do is to use CPPN's to produce images of much higher resolution.

Open Source Agenda is not affiliated with "Logograms GAN" Project. README Source: platers/Logograms-GAN
Open Issues
Last Commit
5 years ago

Open Source Agenda Badge

Open Source Agenda Rating