Artificial intelligences have tried to recognize objects, paint paintings, write films, and of course, generate human faces. In many cases the results were horrible, but time is on their side, and with each new attempt they do a little better. Today it's Nvidia's turn, which just published information about a special training method for generative adversarial networks. While it's not free of errors, the quality of some images is far above what we expected, to the point that they could easily be confused with real people…
30 years ago we saw Mega Man blink for the first time and were amazed. Today, a 'modern' game that features characters with slightly deformed faces and poor animations can be condemned overnight, just as happened to Mass Effect: Andromeda. Digitization and motion capture have become critical resources for many developers, reaching a high level of precision, but out there we can find very big efforts that seek to put that responsibility on artificial intelligences. A solid example is Nvidia, which shared a study on advanced training methods for the so-called generative adversarial networks.
In simple terms, what Nvidia did was train the generator network and the discriminator network progressively, starting with a low resolution and adding new layers that provide much finer details as training advances. The first benefit translates into greater speed, but that pales next to the quality of the images obtained. Not a single person you will see below is real. The trained eye will be able to find several errors without too much effort, however, we must not forget that this is only the beginning.
The gallery
And to finish, the official video that shows the transition and processing capability of the neural network. As Nvidia does, we recommend playing it in 1080p60. It's excellent.
