Core Insights - The article discusses the launch of a new 3D world generation model called Marble, developed by Fei-Fei Li's World Lab, which allows users to easily create personalized 3D worlds without needing a professional team [3][5][15]. Group 1: Model Features - Marble enables users to generate 3D worlds using simple text prompts, single images, or even short videos, making it accessible to the general public [5][17]. - The model includes built-in AI editing tools that allow users to make both minor and major modifications to their created worlds, such as removing objects or changing visual styles [21][25]. - Users can export their created worlds in two formats: high-fidelity Gaussian point clouds for rendering in browsers and triangle meshes for compatibility with various industry-standard tools [29][40]. Group 2: User Experience - The model has received positive feedback for its ease of use, with users quickly sharing their creations online [8][15]. - Marble supports multi-modal input, allowing for a variety of ways to create and edit 3D environments, which enhances user engagement and creativity [34][35]. Group 3: Future Developments - The team plans to focus on enhancing interactivity in future iterations of Marble, enabling real-time interactions within the created 3D worlds [36][37]. - The article emphasizes that Marble is a significant step towards achieving a "truly spatially intelligent world model," which will incorporate capabilities for dynamic interaction and evolution over time [40].

闻乐 发自 凹非寺 量子位 | 公众号 QbitAI 该模型叫 Marble ，是李飞飞创立的World Lab推出的全新3D世界生成模型，直接给世界模型赛道按下加速键——人人都能轻松get专属3D 世界。 就在今天，李飞飞发布了全新的世界模型，开启公测，人人可玩。 创造力就是有趣的智能。 不需要专业团队建模，普通人可以靠文本、照片甚至短视频，轻松生成可编辑、可下载的专属3D世界。 还能在VR里沉浸式体验 ！ 好家伙，就这么一会儿，大家已经带着自己的3D世界刷屏了。 网友已经玩疯了 一键从像素世界"穿越"到彩色艺术小巷，街道是砖石铺就，两侧的店铺门口摆放着黑板菜单牌，还有一辆自行车停靠在路边，很悠闲的市井氛 围～ 或者是未来感和复古感交织的赛博朋克世界。 还内置了AI原生世界编辑工具。编辑可以很小很局部，比如移除一个物体，修饰一个区域。也可以很彻底：交换物体，改变视觉风格，或者重 构世界的大片区域。 带有机械元素的睡眠舱。 充满自然气息的欧式风格庭院。 总之，Marble一经开放就大获好评：使用非常简单。 在官方的技术报告中，Marble不仅能通过简短的文本提示、单图提示生成3D世界，还能通过多张图片、不同视 ...

Core Viewpoint - Tencent's HunyuanWorld 1.0 model enables the generation of immersive 3D worlds from simple text or images, offering high-quality outputs with low operational barriers and compatibility with traditional CG pipelines [5][41]. Technical Framework - The core technology of HunyuanWorld 1.0 utilizes panoramic images as a bridge for layered 3D generation, leveraging the diversity of 2D generation techniques to create rich scenes [9]. - The scene generation process involves three key steps: creating a seamless 360° panoramic image from input text or images, breaking the panoramic image into independent semantic layers, and converting these layers into 3D structures with depth annotations [11][15][16]. Optimization Techniques - The model incorporates two practical optimizations: seamless roaming for long-distance scenes using point cloud caching and video diffusion technology, and dual-mode compression for online/offline storage and inference of 3D models [18]. - Initial versions required over 26GB of VRAM, limiting accessibility for most consumer-grade graphics cards [19]. The introduction of HunyuanWorld 1.0-Lite allows operation on consumer-grade GPUs by optimizing memory usage through dynamic FP8 quantization, reducing VRAM requirements by 35% to below 17GB [20][25]. Performance Enhancements - Dynamic FP8 quantization adjusts the quantization range based on parameter distribution, maintaining model performance while reducing memory usage [26]. - SageAttention quantization technology enhances inference speed by over 2 times with less than 1% precision loss, significantly lowering the memory required for model operation [28][29]. - The integration of a Cache algorithm improves inference efficiency by optimizing redundant time steps, resulting in smoother model operation [33]. Comparative Analysis - HunyuanWorld 1.0 outperforms other open-source 3D models in clarity, inference speed, compatibility with 3D engines, and editability [38]. - It generates editable 3D mesh files rather than videos, making it more versatile compared to competitors like Google's Genie3 [41]. - The model's compatibility with existing CG and 3D production pipelines enhances its practical value, while its open-source nature and single-card deployment capability facilitate easier implementation compared to other models [42].

Core Insights - Tencent officially launched and open-sourced the industry's first 3D world generation model, named Hunyuan 3D World Model 1.0, allowing users to create navigable 3D worlds in minutes by inputting a sentence or an image [1] Group 1 - The new model significantly reduces production cycles by enabling the output of standardized 3D assets [1] - Tencent plans to open-source many more models in the future, including edge-side mixed inference large language models and multimodal understanding models [1]