“Paint using neurons instead of pixels,” promises GAN Dissection a framework to “let you explore what a GAN (generative adversarial networks) has learned by examining and manipulating its internal neurons.” Check out how it can invent details like trees & doorways based on the target image:
I’m reminded of the O.G. PatchMatch demo from 10 (!) years ago that led us to put Content-Aware Fill (itself based on a subset of that work) into Photoshop: