Microsoft has teamed up with Teachable AI Experiences and Xbox Game Studios’ Ninja Theory to introduce a cutting-edge AI model designed to assist game developers. The new model is called Muse.
What is Muse Capable of?
Muse has the ability to generate game visuals and controller actions by interpreting a 3D understanding of a game. The model was trained on 1.6 billion parameters, utilizing a dataset of 1 billion images and controller actions—equivalent to about 7 years of continuous human gameplay. The training data primarily came from Ninja Theory’s Bleeding Edge game.
The data was gathered with the consent of users who contributed gameplay data. The model processes visuals and recorded controller actions at a resolution of 300 x 180px. As a result, Muse can generate up to 2-minute sequences and predict the future evolution of the game in a world model mode. Demos have shown that the model can effectively handle camera shifts, path choices, and even the creation of new characters, all while adhering to the game’s physics.
>>>BTE-002 for Fukuda FX-8322/R FX-8600 FCP-8321 FCP-8453 510114040
Microsoft has already begun using Muse to develop real-time playable AI models trained in other games. This capability allows developers to experiment with new concepts and perhaps even expand on existing titles.
An additional exciting possibility is the implementation of retro games for old hardware. The AI model could be used to adapt classic, vintage games to work on modern machines.
Muse’s model weights, sample data, and the World and Human Action Model (WHAM) are all accessible via Azure AI Foundry. Additionally, interactive AI game demos will be available through Copilot Labs later this year.