Called GAN Theft Auto, this can be GTA V since you haven’t seen it before.

Artificial intelligence figures prominently in a wide swath of games, mostly in how NPCs react to situations. However, AI is really capable of much more than with a personality duck behind a corner in a shootout and, going in a very different direction, a few AI researchers tapped right into Nvidia’s GameGAN neural network to make GAN Theft Auto–a completely AI-generated version of Grand Theft Auto V. The outcome is pretty remarkable.

The playable demo is made up of driving down a brief stretch of highway from GTA V. From a modern graphics standpoint, it is not exactly what you’d expect–that the scene is extremely pixelated, and even after upsampling the outcome signal, there is still a haze, as if playing GTA V at a dream state.

Even so, there are a few notable details visible from the presentation, like generating a shadow beneath the vehicle and accurate sunlight reflections from the rear window which change as the vehicle moves around.

GameGAN is special generative adversarial network (or GAN) created by Nvidia, thus the name GAN Theft Auto. Since the name kind of implies, GANs are made up of two neural networks which are adversaries, in a sense. Among them is them is that a generator. A generator is trained using specific data, for example photos of humans, and then takes what it learns to produce fake content that can pass as the real thing.

The second part is a discriminator. This part is tasked with maintaining the generator fair, so to speak, by discerning what’s real from what is fake.

“GANs can produce images that look like photos of human faces, though the faces do not belong to any real individual,” Nvidia explains. “GANs achieve this level of realism by pairing a generator, which learns to create the target outputsignal, with a discriminator, which learns to distinguish accurate data from the output of this generator. The generator attempts to fool the discriminator, and the discriminator attempts to keep from being duped.”

It was about this time last year after Nvidia introduced GameGAN, the first neural network model that mimics a pc engine using a GAN, also showcased it cloning a completely functional version of Pac-Man. Now it has been used to recreate GTA V.

Kinsley and collaborator Daniel Kukiela made a DGX Station A100 from Nvidia for this project, which was equipped with four A100 Ampere GPUs plus also a 64-core AMD Epyc chip.