ActionParty: Multi-Subject Action Binding in Generative Video Games

TL;DR

ActionParty is a novel multi-subject world model for generative video games that enables simultaneous control of multiple agents with improved accuracy and identity consistency.

Contribution

It introduces subject state tokens and a spatial biasing mechanism to disentangle global video rendering from individual subject actions, allowing multi-agent control.

Findings

Controlled up to seven players across 46 environments.

Achieved significant improvements in action-following accuracy.

Enabled robust autoregressive tracking of subjects.

Abstract

Recent advances in video diffusion have enabled the development of "world models" capable of simulating interactive environments. However, these models are largely restricted to single-agent settings, failing to control multiple agents simultaneously in a scene. In this work, we tackle a fundamental issue of action binding in existing video diffusion models, which struggle to associate specific actions with their corresponding subjects. For this purpose, we propose ActionParty, an action controllable multi-subject world model for generative video games. It introduces subject state tokens, i.e. latent variables that persistently capture the state of each subject in the scene. By jointly modeling state tokens and video latents with a spatial biasing mechanism, we disentangle global video frame rendering from individual action-controlled subject updates. We evaluate ActionParty on the Melting Pot benchmark, demonstrating the first video world model capable of controlling up to seven players simultaneously across 46 diverse environments. Our results show significant improvements in action-following accuracy and identity consistency, while enabling robust autoregressive tracking of subjects through complex interactions.