Abstract

GANs can generate photo-realistic images from the domain of their trainingdata. However, those wanting to use them for creative purposes often want togenerate imagery from a truly novel domain, a task which GANs are inherentlyunable to do. It is also desirable to have a level of control so that there isa degree of artistic direction rather than purely curation of random results.Here we present a method for interpolating between generative models of theStyleGAN architecture in a resolution dependant manner. This allows us togenerate images from an entirely novel domain and do this with a degree ofcontrol over the nature of the output.

 

Quick Read (beta)