今年7月，腾讯混元团队发布混元3D世界模型1.0，支持文本或单张图片输入生成兼容图形学管线的3D 场景；10月，混元团队发布世界模型1.1，支持多视图或视频一键创造3D世界。此次发布的混元世界模 型1.5进一步带来了世界建模的全新可能性。 同时，混元世界模型1.5(WorldPlay)首次开源了业界最系统、最全面的实时世界模型框架，涵盖数据、 训练、流式推理部署等全链路、全环节，并提出了重构记忆力、长上下文蒸馏、基于3D的自回归扩散 模型强化学习等算法模块。 12月17日，腾讯混元发布世界模型1.5(Tencent HY WorldPlay)，用户输入文字描述或者图片即可创建专 属的互动世界，通过键盘、鼠标或手柄实时控制虚拟相机的移动和转向，像玩游戏一样自由探索AI生 成的世界。这是国内首个开放体验的实时世界模型。 ...

Core Viewpoint - Enigma Labs has developed the world's first AI-generated multiplayer game, Multiverse, which allows players to interact in a dynamically evolving world at a low development cost of under $1,500 [2][3]. Group 1: Game Development and Features - Multiverse is a multiplayer racing game where players can overtake, drift, and accelerate, reshaping the game world with each action [2][3]. - The game operates on a model that allows real-time interaction with an AI-simulated environment, filling a gap in AI-generated worlds [3][6]. - The development team plans to open-source all related research, including code, data, and architecture [3][8]. Group 2: Team Background - The team consists of former members of Israel's elite 8200 unit and experienced professionals from leading startups, specializing in research and engineering [5]. Group 3: Technical Architecture - The architecture of the multiplayer model builds upon existing single-player models, requiring a redesign of input and output connections to facilitate cooperative gameplay [12][14]. - The model integrates player actions and frame data to ensure a consistent shared world state, crucial for multiplayer interactions [14][15]. Group 4: Data Collection and Training - The training data for the model was sourced from Sony's Gran Turismo 4, with the team modifying the game to enable a 1v1 mode for data collection [39][41]. - The team utilized computer vision to extract control inputs from HUD elements displayed during gameplay, allowing for the reconstruction of player actions without direct input recording [46]. - A scalable method for data generation was implemented using the B-Spec mode, enabling automated race recordings from multiple perspectives [48]. Group 5: Model Training and Performance - The model was trained to predict future frames over varying time horizons, initially focusing on short-term predictions before extending to longer-term interactions [32][33]. - Efficient long-range training techniques were developed to manage memory constraints while maintaining performance [34][35].