GameNGen

Google Research

Yaniv Leviathan*

Google Research

Moab Arar*†

Tel Aviv University

Shlomi Fruchter*

Google DeepMind

• Equal Contribution Work done while at Google Research

†

Paper Arxiv

Real-time recordings of people playing the game DOOM simulated entirely by the GameNGen neural model.

Abstract

We present GameNGen, the first game engine powered entirely by a neural model that enables real-time interaction with a complex environment over long trajectories at high quality. GameNGen can interactively simulate the classic game DOOM at over 20 frames per second on a single TPU. Next frame prediction achieves a PSNR of 29.4, comparable to lossy JPEG compression. Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation. GameNGen is trained in two phases: (1) an RL-agent learns to play the game and the training sessions are recorded, and (2) a diffusion model is trained to produce the next frame, conditioned on the sequence of past frames and actions. Conditioning augmentations enable stable auto-regressive generation over long trajectories.

Full Gameplay Videos

Architecture

Architecture overview

Data Collection via Agent Play: Since we cannot collect human gameplay at scale, as a first stage we train an automatic RL-agent to play the game, persisting it's training episodes of actions and observations, which become the training data for our generative model.

Training the Generative Diffusion Model: We re-purpose a small diffusion model, Stable Diffusion v1.4, and condition it on a sequence of previous actions and observations (frames). To mitigate auto-regressive drift during inference, we corrupt context frames by adding Gaussian noise to encoded frames during training. This allows the network to correct information sampled in previous frames, and we found it to be critical for preserving visual stability over long time periods.

Latent Decoder Fine-Tuning: The pre-trained auto-encoder of Stable Diffusion v1.4, which compresses 8x8 pixel patches into 4 latent channels, results in meaningful artifacts when predicting game frames, which affect small details and particularly the bottom bar HUD. To leverage the pre-trained knowledge while improving image quality, we train just the decoder of the latent auto-encoder using an MSE loss computed against the target frame pixels.

BibTeX

@misc{valevski2024diffusionmodelsrealtimegame,
        title={Diffusion Models Are Real-Time Game Engines},
        author={Dani Valevski and Yaniv Leviathan and Moab Arar and Shlomi Fruchter},
        year={2024},
        eprint={2408.14837},
        archivePrefix={arXiv},
        primaryClass={cs.LG},
        url={https://arxiv.org/abs/2408.14837},
  }

Acknowledgements

We'd like to extend a huge thank you to Eyal Segalis, Eyal Molad, Matan Kalman, Nataniel Ruiz, Amir Hertz, Matan Cohen, Yossi Matias, Yael Pritch, Danny Lumen, Valerie Nygaard, the Theta Labs and Google Research teams, and our families for insightful feedback, ideas, suggestions, and support.