Generative Models as Data++

Phillip Isola

Massachusetts Institute of Technology, USA

Abstract

This lecture will view generative models as a new kind of interface to data. I will cover the basics of popular deep generative models — GANs, autoregressive models, diffusion models — and will show how these models can be steered to create “living”, controllable data as an output. I will end by envisioning a future in which datasets have been replaced by models, and discuss how samples from these models can be used to train downstream vision systems.