Gamma-World
Implementation of Gamma-World: Generative Multi-Agent World Modeling Beyond Two Players
Gamma-World is a research project from NVIDIA and academic collaborators that generates interactive video environments where multiple independent agents, such as game players or robots, can act simultaneously and see a consistent shared world. The paper and project page were released in May 2026, but the code has not been published yet.
The core problem it addresses is that most AI video generation systems are designed for a single character or viewpoint. When you have two or more agents each making their own decisions, the AI needs to keep the world consistent for all of them at once. Each agent must be independently controllable, and swapping the order of agents in the input should not change how the model treats them. Gamma-World is designed to handle all of this while staying fast enough to generate video in real time.
To solve the agent identity problem without using a fixed ordering, the model places each agent at a vertex of a geometric shape called a regular simplex in a mathematical space used for position encoding. Every pair of agents is equally spaced, so no agent is treated as special or first, but each still has a distinct identity the model can track.
For efficiency, the model avoids calculating direct interactions between every possible pair of agents, which would become very expensive as the number of agents grows. Instead, a small set of learnable intermediate tokens called hub tokens collect information from all agents and redistribute it. This keeps the cost growing linearly with the number of agents rather than as a square.
The result runs at 24 frames per second. A model trained on two-player scenarios can generalize to four players without additional training, and the system has been tested on both video game environments and real-world robot coordination tasks.
As of the repository creation date, the code and trained model weights have not been released. The repository currently contains the paper, example images, and a project overview. The authors have indicated that code and checkpoints are planned for a future release.