// the find
Sentdex/GANTheftAuto
A fork of Nvidia's GameGAN that attempts to run GTA5 inside a neural network — the GAN learns to generate plausible next frames given player input, so there's no game engine at inference time. It's research code, not a library, aimed at ML researchers curious about neural game simulators or anyone who wants a starting point for applying GameGAN to their own environment.
The sample GTA5 dataset and pretrained model are included, so you can run inference without owning a copy of GTA5 or setting up the data collection pipeline. The addition of a Vroom environment (no OpenAI Gym dependency, just OpenCV and NumPy) makes it easier to experiment with a simpler domain first. Support for non-square images and more than 2 generators are real improvements over the upstream GameGAN code, not just cosmetic changes. TensorBoard integration is wired in, so training progress is observable without extra setup.
The data collection scripts for GTA5 are withheld — you get a small sample dataset but can't collect more without reverse-engineering the approach or begging the authors. Inference is explicitly described as unfinished in the README (actions are random, not player-driven), which means the 'playable demo' claim is mostly aspirational. Last commit was 2022, PyTorch has moved significantly since 1.8.1, and there's no indication anyone has tested compatibility with current versions. The Cartpole environment has a known bug the authors flagged and never fixed.