face_collage jpg
(68.93 KB, 876x257)
Here's some tech that looks like it's poised to turn commercial image creation upside down. It's easy to imagine a number of highly desirable image generation products built around this, including texture generation for games and serious competition for stock photo sites.
The faces shown are all artificially generated. The same thing can be done with any object that enough images are available of. Want rocks, cars, buildings, landscapes? It looks like this approach could be used to make a random image generator for any particular thing you wanted.
With your standard AI, you set the weights of nodes in a neural net with a bunch of images that you know classifications of (male vs female, cat vs dog, etc). You keep at it until the net starts returning a ratio that you're happy with of correct-to-incorrect guesses when presented with an image that it hasn't seen before. The goal is to build a machine that can successfully say what something is.
With this Adversarial Network stuff, you train the first network a little differently: Instead of deciding what an image is, its job is to decide whether the input image is from the training set or not. A second network is also trained up, and its job is to generate output images that succeed in fooling the first net. These images are the output of that second network.
Cool vid showing image morphs from training:
https://youtube.com/watch?v=XOxxPcy5Gr4
Paper on GANs:
https://arxiv.org/abs/1406.2661
pix2pix and iGAN githubs linked from here:
https://people.eecs.berkeley.edu/~junyanz/