Source: Deep Learning on Medium
The Witcher, painted by a deep neural network
The Witcher has recently been released on Netflix. Originally a book series by Polish author Andrzey Sapkowski, the story about a monster-hunting mutant reached a wide audience when developer CD Projekt turned it into a video game series. The games meant that, when Henry Cavill was cast as the protagonist Gerald of Rivia, fans already had an idea of what their favorite monster hunter looked like.
I wondered — what would video-game Gerald’s face be like if not rendered by a video game engine, but painted by the generative deep neural network StyleGAN developed, trained, and released by Nvidia? This network has been trained to turn random numbers into photo-realistic “photos” of imaginary people, but it is, in principle, powerful enough to create images that appear just like specific people, if presented with the right inputs. By feeding an image into yet another network and optimizing for a while, we can determine the right input to recreate said image. This works well for most ordinary photos, but we can also try our luck with a screenshot of a video game character — such as Gerald, the Butcher of Blaviken.
Not bad, even though it is not really photo-realistic, and it lacks the scars video-game Gerald carries on his face. Well, so does Henry Cavill. Just for fun, let us process and recreate a photo of him as the Witcher.
Close enough to the actor, as well as the character. Even though StyleGAN tends to draw overly smooth features, it captures a lot of the grim expressiveness and has no problems with the extreme lighting.
Now, to wrap things up, let us try one of Gerald’s companions, Triss Merigold. She is known for her flaming red hair, which could present a problem for StyleGAN. Here goes…