Recently, Kunlun AI officially announced the open-source release of its latest Matrix-Game large language model. With a parameter scale exceeding one billion, this model is the first industrial space intelligence large language model to be open-sourced, marking a significant breakthrough in interactive world generation technology. Matrix-Game not only supports the well-known game "Minecraft" but is also specifically designed for high-quality generation and precise control in open environments.
The core of the Matrix-Game large language model lies in its three main components. First is the "Matrix-Game-MC dataset," which is self-constructed and includes a large number of Minecraft game videos, such as unlabeled large-scale videos and controllable video data with control signals. This allows developers to efficiently model dynamic and interactive patterns in complex environments. Second, the main model of Matrix-Game utilizes advanced diffusion model technology to generate coherent and controllable interactive videos based on user input (such as keyboard and mouse operations), balancing visual effects, temporal consistency, and physical plausibility. This means players can experience more realistic interactions in games.
Source Note: Image generated by AI, image authorized service provider MidJourney
Finally, Matrix-Game introduces the GameWorld Score evaluation system, a new standard for evaluating game interaction worlds. It comprehensively quantifies model performance from multiple dimensions such as visual quality, temporal quality, action controllability, and understanding of physical rules, filling the gap in systematic evaluation benchmarks in this field. This evaluation system will help developers better understand the advantages and disadvantages of the model.
Matrix-Game can achieve controllable generation in various Minecraft scenarios, supporting dynamic behaviors of characters in desert, forest, and other environments. Users can experience character movements, jumps, and attacks through simple control instructions. Moreover, this model supports autoregressive long video generation, ensuring seamless transitions between actions and perspectives, laying a solid foundation for immersive experiences and creative content generation.
Kunlun AI's Matrix-Game large language model is not only a technical innovation but also a milestone in game development, and we look forward to its extensive application in the future.