TL;DR: A couple of programmers wanted to see if they could train a generative adversarial network (GAN) to create its own version of Yard Theft Car 5. The result was a somewhat blurry acid trip but is instantly recognizable as GTA5.

Harrison Kinsley, who goes by Sentdex on YouTube, and his partner trained a fork of Nvidia'southward GameGAN neural network using a black motorcar on a brusk section of highway in Grand Theft Auto 5. Nvidia loaned them its DGX Station, which is equipped with four A100 80-gigabyte cards to aid with the processing. Kinsley explains what they did and demos the results in the video above.

Initial models were very pixelated, simply Kinsley improved this with AI-assisted supersampling. While it'due south not pretty, it's important to go on in heed that this is non GAN-generated footage. Information technology is a real-time interactive demo. Kinsley is driving an AI-created car in a fully AI-created environment.


Before and afterward upsampling.

Kinsley said that since the DGX Station was on loan, his training fourth dimension was limited. Although the model does attempt to compute obstacle clipping in some instances, he would have liked to run more collision samples. Kinsley likewise wanted to run across how much of the GTA5 map the GAN could process. However, that would take required hours more than grooming equally he incrementally increased the driving distance, which he did non have time to practise.

The video shows the neural network did a pretty decent job recreating some unexpected details. For efficiency'due south sake, one might retrieve that the GAN would ignore shadows and the sun reflecting off the auto. To Kinsley'due south surprise, it didn't. Shadows, lighting, and reflections move more than or less every bit expected. The GAN besides created its own elementary physics system after training it by running into things. An instance is how some other car skews to the right when tapped on the left rear quarter panel. Head-on collisions are not handled also.

Kinsley has posted the playable demo dubbed "GAN Theft Auto" to GitHub for those interested in giving it a spin. Again, it's only a tiny portion of the GTA map, and information technology's not like playing a real game. It'due south more than tech demo than anything else, simply it is interesting to watch what the model does in untrained situations. That'due south when things become a trivial psychedelic.