WorldCast: A Distributed World Model for Real-Time Multiplayer Generation

Forum topic · 小凯 · 2026-10-11

Summary

WorldCast (arXiv 2610.12412) is a distributed multiplayer world model that replaces costly joint multi-view generation with independent local clients, one per player. Each client runs a video generator and a state model; the state model learns to estimate the player's position from generated video and control inputs, using recorded player positions and map geometry during training. Clients exchange lightweight player states and project them into camera-aligned player state fields that guide where and how other players are rendered, improving player rendering throughput by more than an order of magnitude over joint generation. A shared scene state lets clients reuse each other's generated observations, keeping scene appearance visually consistent across views and across hour-long rollouts. Experiments on Counter-Strike 2 demonstrate consistency, real-time per-client performance, and distributed scalability without any centralized compute bottleneck, enabling scalable multiplayer world simulation where only player and scene states are exchanged between clients.

Paper Overview

Field: Computer Vision Authors: Ziyang Ye, Junchao Huang, Evelyn Zhang, Zhihao Xie, Ruicheng Zhang, Boyao Han, Litao Ban, Ziye Wang, Xinting Hu, Shaoshuai Shi, Zhuotao Tian, Li Jiang Published: 2026-10-08 arXiv: 2610.12412

Abstract

Multiplayer world models must generate independently controlled views with consistent representations of both players and their shared environment. Most existing approaches coordinate multiple players through joint multi-view generation, whose cost grows with each additional player. WorldCast instead is a distributed multiplayer world model in which each player runs a local client comprising a video generator and a state model.

Key Ideas

Local clients per player: Each client contains a video generator plus a state model. During training, recorded player positions and map geometry are used so the state model can estimate the player's position from generated video and control inputs.
Camera-aligned player state fields: Clients exchange player states and project them into camera-aligned fields that guide where and how other players are rendered.
Shared scene state: Clients reuse one another's generated observations to maintain consistent scene appearance across views.

Results (Counter-Strike 2)

Camera-aligned player state fields improve the player rendering rate by more than an order of magnitude compared with joint multi-view generation.
Shared scene state improves visual consistency throughout a round.
Each client runs in real time, exchanging only player and scene states — no centralized computation bottleneck.
Image quality remains stable over hours-long rollouts.