Controllable content generation

C├ędric Beaulac, Jeffrey S. Rosenthal and David K. Duvenaud

Submitted to Machine Learning for Creativity and Design workshop at NeurIPS 2019. (Rejected)

Abstract :

We use the recent advances in generative models to a construct controllable content generator that extends the current procedural content generator abilities to democratize artistic content creation. Our goal is to give content creators the power to decide the properties of the content they would like to control, to give them tangible control over these properties and let generative models fill the gaps. Our main contribution is to define what control is and how to measure it. We also propose a model based on the Variation AutoEncoder architecture that is inspired from semi-supervised learning to create a controllable content generator.

Download here