# Gamma-World

Gamma-World is a generative multi-agent world model that generates coherent future video frames for multiple independently controllable agents acting in one shared environment, scaling from multiplayer games to multi-robot settings.

Gamma-World is the code and training pipeline for a generative multi-agent world model that interactively generates future video frames from the combined actions of multiple agents while keeping the shared environment consistent across time and viewpoints. It introduces Simplex Rotary Agent Encoding, a parameter-free extension of 3D RoPE that places agents as vertices of a regular simplex in rotary angle space so each agent gets a distinct phase while all agents remain interchangeable, and Sparse Hub Attention for efficient communication between agents.

It streams generated video in real time at 24 frames per second using a distilled block-causal student model, and generalizes to more players than it was trained on, for example from two to four, without retraining. The repository includes setup instructions, inference on bidirectional, causal and causal-few-step model variants from a user's own initial frames and actions, and training scripts that convert a released checkpoint and train the teacher, causal and DMD stages.

It is written in Python, developed by NVIDIA with Tsinghua University, the University of Toronto and the Vector Institute, and released under Apache-2.0.

Website: https://research.nvidia.com/labs/sil/projects/gamma-world/