Neura Market
GameNGen logo

GameNGen

Free

A game engine powered entirely by AI, capable of simulating DOOM in real time. Opens new perspectives for game development

4.4
GamesFreeFree tier
#game-engine#research
Inputs: image, textOutputs: image
Type
Saas
Visit Website

About GameNGen

GameNGen is a research project from Google Research and Tel Aviv University that introduces a neural model capable of serving as a real-time game engine. The model, built on a diffusion architecture, interactively simulates the classic game DOOM at over 20 frames per second on a single TPU, with next-frame prediction achieving a PSNR of 29.4, comparable to lossy JPEG compression. Human evaluators are only slightly better than random at distinguishing short clips of the simulation from the actual game. The development of GameNGen involves two phases: first, an RL agent learns to play the game and its training sessions are recorded; second, a diffusion model (based on Stable Diffusion v1.4) is trained to produce the next frame conditioned on past frames and actions. Conditioning augmentations and latent decoder fine-tuning help maintain stable auto-regressive generation over long trajectories. As a purely research-oriented demonstration, GameNGen is not a commercial product or publicly available service. Its primary contribution is to show that neural models can act as game engines, potentially opening new avenues for game development and AI-driven simulation. The project page includes the research paper, a bibliography entry, and acknowledgments, but no direct access to a playable or downloadable tool beyond the provided examples and videos.

Key Features

Real-time neural simulation of a complex 3D environment (DOOM) at over 20 fps on a TPU
Uses a diffusion model trained on gameplay data collected by an RL agent
Conditioned on sequences of past frames and actions for auto-regressive generation
Latent decoder fine-tuning improves visual quality, especially for HUD elements
Human evaluators find the simulation nearly indistinguishable from the original game
Demonstrates the first instance of a neural model acting as a complete game engine

Pros & Cons

Pros
  • Breakthrough demonstration of a fully neural real-time game engine
  • High-fidelity simulation with near-human indistinguishability in short clips
  • Open access research with published paper and code (likely available on GitHub)
  • Novel two-phase training approach (RL agent + diffusion model) that can be adapted
Cons
  • Currently only demonstrated on a single game (DOOM); generalizability is not proven
  • Requires high computational resources (TPU) to achieve real-time performance
  • Not available as a commercial product or public service – available for research use only
  • The simulation may exhibit drift over very long trajectories without context corruption methods
  • Potential artifacts from the latent auto-encoder, especially in HUD details (partially addressed)

Best For

Research into AI-driven game engines and real-time simulationExploring the replacement of traditional game engine pipelines with neural modelsGenerating interactive gameplay footage for training or evaluationStudying long-term stability and drift in auto-regressive video generationDeveloping new methods for neural rendering and world models

Alternatives to GameNGen

G3DAI {Jedi}

Rapidly develop game 3D models, generate detailed landscapes from text, and craft exceptional interior design assets.

ImgenArena.ai

A fun platform for evaluating the quality of AI-generated images. Vote for the best match between prompts and images

Joypanner

Encourages children to explore with tailored AI activities and games for problem-solving and creativity.

Chatmate AI

Chat directly with AIs that look very human in your messaging applications

Hidden Door

An AI-driven RPG adventure game that generates creative stories

NVIDIA RTX Remix

Remaster retro games with ray tracing and AI-enhanced textures effortlessly

FAQ

Is GameNGen publicly available for use?
Based on available information, GameNGen is a research project. The website provides a paper and likely code for research purposes, but a ready-to-use simulator or API is not confirmed. Direct playability should be verified on the project’s official page.
Can I play DOOM using GameNGen?
The project demonstrates real-time simulation of DOOM, but it appears to be a research prototype rather than a polished game. Whether the simulation can be interacted with as a game is not explicitly stated; the examples show recorded gameplay, not interactive play.
How does GameNGen achieve real-time performance?
The model runs at over 20 frames per second on a single TPU. It uses a diffusion model trained on RL agent data, with conditioning on past frames and actions. Performance on consumer hardware is not addressed.
What hardware is needed to run GameNGen?
The reported performance is on a TPU. The paper does not specify requirements for other hardware. Running on GPUs or CPUs may not achieve real-time speed.
Is GameNGen capable of simulating other games?
The research only demonstrates simulation of DOOM. The method could potentially be adapted to other games, but no evidence of such generalization is provided in the current project.