MultiGen:可编辑多人世界的关卡设计在扩散游戏引擎中的应用

📄 中文摘要

视频世界模型在互动模拟和娱乐领域展现了巨大的潜力,但当前系统在用户对环境的控制和共享推理方面仍存在不足。为了解决这些问题,引入了一种显式的外部记忆,作为独立于模型上下文窗口的持久状态,能够持续更新用户的操作并在生成过程中进行查询。与传统的扩散游戏引擎不同,该方法将生成过程分解为记忆、观察和动态模块。这种设计使得用户能够在一个共同的世界中进行可重复和可编辑的体验,从而提升了互动性和用户参与感。

📄 English Summary

MultiGen: Level-Design for Editable Multiplayer Worlds in Diffusion Game Engines

Video world models have shown significant potential in interactive simulation and entertainment; however, current systems still face challenges in user control over the environment for reproducible and editable experiences, as well as shared inference where players can influence a common world. To address these issues, an explicit external memory is introduced, functioning as a persistent state independent of the model's context window, continuously updated by user actions and queried during the generation rollout. Unlike conventional diffusion game engines that predict the next frame, this approach decomposes the generation process into Memory, Observation, and Dynamics modules. This design enhances interactivity and user engagement by allowing users to create repeatable and editable experiences within a shared world.

Powered by Cloudflare Workers + Payload CMS + Claude 3.5

数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等