ONB Postcards StyleGAN

During the ONB Labs Symposium 2019 Gene Kogan and Sofia Crespo held a workshop helping artists expand their knowledge of technology and helping tech people learn to be more artistic and creative. Wherever participants were on that spectrum, they were invited you to join this workshop on AI art using the ONB Labs data. The course introduced a family of machine learning-based techniques which synthesise, transfer, collage, and remix the styles of images. One of these techniques are Generative Adversarial Networks (GANs), in which two adverse networks operate, one of which is generating results while the other compares it to a set of "real" data. The goal of the generating network is to create results that are indistinguishable from the underlying data. After the ONB Labs event Gene and Sofia trained a StyleGAN based on our historical postcard set. The result presented here in the form of a video is – to say it with the creators words – surprisingly nice.

Using the software introduced during the ONB Labs Symposium Runway application (runwayml.com) Gene and Sofia trained a SyleGAN on the ONB Labs postcard data set. This is their result:

Tools

Tools used to generate visuals

DescriptionLink

RunwayML

Machine learning platform for creators

https://runwayml.com/

Data sets

Used data sets for StyleGAN training

DescriptionLink

Historical Postcards

The dataset contains front images of historical postcards from the early 20th century

Link

Downloads

Further resources for this topic

DescriptionLink

GitLab repository

A repository including the generated video

Link