one year on
Google researchers unveil GameNGen, a diffusion model that simulates DOOM in real time
The first neural game engine runs at over 20 fps on a single TPU, with human raters barely able to distinguish short clips from the real game.
Google researchers have published GameNGen, the first game engine powered entirely by a neural model, capable of simulating DOOM in real time at over 20 frames per second on a single TPU. The system uses a diffusion model — based on Stable Diffusion v1.4 — conditioned on sequences of past frames and player actions. It achieves a PSNR of 29.4 for next-frame prediction, comparable to lossy JPEG compression.
The two-stage training process first has an RL agent play DOOM, recording its sessions. Then a diffusion model is trained on those recordings to predict the next frame, conditioned on past frames and actions. To prevent auto-regressive drift, the researchers corrupt context frames with Gaussian noise during training, allowing the network to correct itself over long trajectories. A fine-tuned latent decoder improves HUD details.
The paper claims human raters are only slightly better than random chance at distinguishing short clips of the simulation from the real game. The project page hosts full gameplay videos. On Hacker News, the thread has 1149 points and 409 comments, with commenters debating whether the approach could generalize beyond DOOM.
The thread reaches 1149 points and 409 comments, with commenters debating whether the approach could generalize beyond DOOM.
One year later — open only if you can handle spoilers
The paper became a reference point for generative game engines, though no commercial product emerged. By mid-2026, diffusion-based game simulators remain a research novelty, with real-time neural rendering still limited to simple environments.
The Weekly Replay · free by email
This week, one year ago — every Sunday.
One email each Sunday: the week's replayed AI news, with the one-year-later annotations included. Written like it's breaking — dated like it isn't.
Free · double opt-in · unsubscribe anytime · privacy