WHAM: Microsoft's AI That Plays and Predicts Gameplay

WHAM also called "Muse" is Microsoft's AI model built to generate game visuals and predict controller moves either separately or together. It comes from Microsoft’s Game Intelligence group working with the Teachable AI Experiences (TaiX) team and Ninja Theory. They trained it on gameplay data from Ninja Theory’s 4v4 multiplayer game Bleeding Edge—over a billion images and controller inputs packed into more than seven years of straight human gameplay.

• WHAM is an AI that creates game visuals and predicts gameplay
• Developed by Microsoft Game Intelligence TaiX and Ninja Theory
• Trained on Bleeding Edge data with over a billion images and actions
• Uses a VQ-GAN encoder and transformer for smooth game sequences
• Available under Microsoft Research License on Hugging Face and Azure AI Foundry
• Could help improve classic games for modern platforms
• Microsoft made a WHAM Demonstrator for hands-on testing

One cool use? WHAM might help keep and improve older games. By looking at past gameplay it could tweak and optimize classic titles making them work better on today’s platforms.

How WHAM Works

WHAM's design has two main parts. First a VQ-GAN encoder-decoder that turns game visuals into smaller building blocks. Second a transformer setup that predicts the next move. This setup lets it build smooth game sequences grabbing the 3D world physics player actions and how the game unfolds over time.

Who Can Use It and Where?

WHAM is available under the Microsoft Research License on platforms like Hugging Face and Azure AI Foundry. Devs and researchers can play around with its model weights and sample data to test what it can do.

Microsoft built a WHAM Demonstrator—a prototype tool with a visual interface where users can try different prompts and see the AI’s output live.

WHAM is a big step for generative AI in gaming letting devs create and predict complex gameplay sequences with a powerful new setup.