The previous version of this published a year ago was using public CelebA dataset; and no, vast majority of generated faces were different than training dataset. You could even change latent variables for continuous "morphs" between various synthetic faces, each stage being a unique face. Read on how GANs work first if you have doubts.

I have a rough idea that how it works but I am still skeptic.

Faces of kids specially are unbelievable to be generated. Their backgrounds, clothes, hair facial expressions seem all very unique to that picture. I will be going through the dataset to find the matches once they release it.

It can be demonstrated that GANs learn to model certain image contents with separate input variables like the presence of certain objects or 3D rotations [0] all by themselves, so there is definitely more going on than simply learning to paint some image patches at certain locations and smoothly interpolating them. An image patch reveals itself directly in the training data, so the image patch could simply be stored in the network weights, but a 3D rotation or the presence of a certain class of objects cannot be learned by simply copying image patches into the weights. A 3D rotation requires at the very least a computation of foreshortening, occlusion perhaps based on a depth map. An object detector requires at the very least a feature hierarchy, perhaps binding computations that relate different object parts to the whole object.

Neural networks basically learn to implement nearly arbitrary computations (up to a certain circuit depth) to produce the desired output, so you can also think of deep learning as program mining of a certain program space that is reachable by the adaptive functions in a neural network. Stephen Wolfram has written about that in his blog and talked about it in one of his podcasts. So it's basically magic much like evolution, and it will probably destroy us because it's simply too powerful.

[0] https://twitter.com/phillip_isola/status/1066567846711476224

I thought it was CelebA-HQ (https://drive.google.com/drive/folders/0B4qLcYyJmiz0TXY1NG02...) but the article states that they have made a better dataset; which isn't published yet but promised to be published in January at https://docs.google.com/document/d/1SDbnM1nxLZNuwD8fQkIigUve...

Per: https://docs.google.com/document/d/1SDbnM1nxLZNuwD8fQkIigUve... They're doing so in January, along with the source code.

Asking the important questions