Stumbling onto Alec Radford’s deep convolutional generative adversarial network, or DCGAN, was one of the few genuinely jaw-dropping moments I’ve experienced in my life.
His imagesof bedrooms were not captured by cameras, but by statistics. They are generated from what you might call the “probability space” of bedroom pictures; the statistical distribution that contains all the various features you would expect to find in a picture of a bedroom, neatly separated so that a random sample from said distribution can be “morphed” into something that resembles a photograph.
The way this is done is both shockingly simple and dreadfully complicated. It’s simple in that it doesn’t require a ton of code, appears to work on various kinds of datasets, and can be run on a medium-grade GPU to start producing interesting results in a few hours.
Read more: Generating Fine Art with 300 Lines of Code