Matrix-Game 4.0 puts generated video over a game engine that keeps the rules

October catch-up. Paper published October 20, 2026. Published November 13, 2026.

Matrix-Game 4.0 gives a conventional game engine responsibility for movement, collisions, and the rules, then asks a video model to redraw its low-poly frames. The October 20 paper describes shooter and driving examples where the underlying game retains its state while the generated visuals change its appearance. Research paper.

The team's original 1080p overview shows the low-poly input, generated appearance, and game events such as shield damage and defeat. Video: Matrix-Game team. These are the researchers' demonstrations and measurements.

Keep a world underneath the pictures

The engine draws a simple version of each frame. The model receives that image, a description of the current scene, and visual references, then generates the display frames in groups of four. A coding agent can change rules and supply descriptions derived from the game's state. The project page includes synchronized comparisons where you can move a divider between the engine's output and the generated image.

This division matters for a playable game. Whether a shield has broken or an enemy has been defeated can stay in executable code. The renderer's job is to depict that state convincingly. Visual consistency still needs testing, especially after long movement through a scene or repeated changes of view.

What the speed claim covers

The paper reports 480p gameplay at 60 FPS on one NVIDIA RTX 5090, with 112 milliseconds from key press to a responding display frame. It separately reports 83 FPS of steady-state model throughput with one denoising step and 18.2 GiB of peak GPU memory. Those figures describe different measurements. They shouldn't be treated as interchangeable estimates of what a remote player will experience. Performance results.

The released implementation supplies code, downloadable scene resources, offline rendering, and a single-user WebRTC browser demo. Inference runs on a Linux NVIDIA GPU host. The browser receives the stream, so the client doesn't need to run the model itself. Encoding, transport, buffering, and decoding add delay beyond inference.

The repository also supplies RTX 5080 and 5070 Ti presets with fixed-prompt restrictions. Its reported throughput excludes video input and output. For a developer trying the release, the useful first comparison is the same scene with and without neural rendering, followed by a measured browser session on the connection the intended player will use. Setup and measurement notes.