Skip to content

WHAM

Generate game visuals and controller actions from prompts

Visit official site

What WHAM is

The model is described as WHAM, short for World and Human Action Model, developed by Microsoft Research for generating game scenes and player behaviour. It is trained on data from Ninja Theory's Bleeding Edge and produces coherent, diverse game visuals and controller actions.

What you can do with it

  • Generate game visuals and controller actions
  • Work in world modelling, behaviour policy and full generation modes
  • Capture the 3D structure of game environments and the time sequence of player actions
  • Generate game sequences from initial visuals or controller actions as prompts

Application scenarios

  • Generate Bleeding Edge character actions and scenes
  • Support creative iteration in game design through model inference
  • Display generated visuals and controller actions in real time with the demo tool

Getting started

  • Clone the WHAM GitHub repository and set up a virtual environment
  • Download the model weights for the 200M or 1.6B parameter model
  • Prepare sample data or use the provided sample data
  • Run the local inference script to generate a game sequence

Pros & cons

✓ What we like

  • Initial visuals or controller actions can serve as prompts
  • Two model sizes are offered for different needs

! What to watch out for

  • Training data is described as coming from one game
  • Generation centres on game visuals and controller actions

FAQ

What does WHAM stand for?

The material expands it as World and Human Action Model.

What model sizes are listed?

The material lists 200M and 1.6B parameter models.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review