Finally, the HD remake we’ve all been waiting for: a series of neural nets took the classic pixelated mug of the space marine from seminal 1993 first-person shooter “Doom” and churned out an almost photorealistic portrait.
The face, which looks sort of like a beefier Nathan Fillion, was posted to Reddit on Sunday. And while the resulting thread reveals that the process wasn’t as straightforward as it seems, the image is still a startling example at how adept artificial intelligence has become at filling in the gaps in a low-fidelity input — and spitting out something new and convincing.
Step By Step
The final image is actually the product of several AI systems taking the improving on one another. According to the Reddit thread, the original “Doom Guy” sprite was fed into the AI of photo-editing programs including FaceApp, Waifu2x, and GIMP. That resulted in an improved, but still very pixelated and maybe pinkeye-infected version of the face.
From there, the image was fed into StyleGAN, the Nvidia AI system that people have used to create photorealistic portraits and nightmarish Pokémon sprites. StyleGAN churned out a more realistic image then took another pass to smooth it out. But because the original sprite had unrealistic proportions, the new image was manually reshaped, into this final version.
The Short Version
Basically, one algorithm couldn’t take the pixelated blob that is Doom Guy and turn it into a real-looking face — it took a few steps to get there.
But once that process is set up, AI can do some cool stuff — the original Reddit post, for instance, shows how editing teeth into the pixelated sprite convinces the AI to create a very different portrait, reshaping Doom Guy’s entire expression into a big beaming smile.
More on StyleGAN: This Site Tests Whether You Can Spot AI-Generated Faces