These are some experiments made using Deep Generator Networks (DGN) code by Nguyen, Dosovitskiy, Yosinski, Brox, and Clune at Evolving-AI. DGNs are a class of neural networks which synthesize near-photorealistic samples from a probability distribution learned over a large collection of images.
All of these were generated using their original code, applied to a caffemodel of Places365. I’ve selected some of the interesting highlights.
DGNs are capable of producing many unique samples which are pretty convincing. Here is a series of buttes.
Here is a series of boathouses. Notice that the DGN has impressively learned to make the water reflective.
Since many image classes have people in them, they have a tendency to generate people as a strange artifact. Curiously reminiscent of Mario Klingemann’s work with DGNs.
At the discotheque.
You can animate the process of DGN samples converging on a class.
watching DGN samples form. they jostle wildly through a generative space until converging on… a cheeseburger pic.twitter.com/yEKsTgMweu— Gene Kogan (@genekogan) August 29, 2016