Matrix Game

Matrix Game 2.0 is an open-source interactive world foundation model from SkyworkAI that generates real-time, long videos at 25 fps in virtual environments. Built on an auto-regressive diffusion-based image-to-world framework, it responds to keyboard and mouse inputs, enabling fine-grained control and dynamic scene evolution. Designed for researchers, game developers, and AI enthusiasts, it provides a platform for interactive world simulation without requiring game engines. The model runs locally or via provided inference scripts, supporting streaming generation. Key differentiators are its real-time performance (25 fps), open-source nature, and direct user control through standard input devices. It leverages advanced diffusion techniques for coherent long-form video generation from a single starting image. Users can clone the repository, install dependencies, and run inference scripts to generate interactive video streams. The project includes demo utilities and configuration files for various scenarios. Three pretrained models are available on Hugging Face: universal scenes, GTA driving scene, and TempleRun game scene. A known limitation is that upward camera movement may cause brief rendering glitches (black screens), with a fix planned.