World Labs 最近推出了 Atlas，这是一款新型的“前沿模型”，它获得了“惊人的反响”。Feifei、Justin 和 Ben 讨论了它的能力、重要性以及未来的影响。Atlas 被描述为一款“下一代世界模型”，能够生成、重建和模拟世界。其核心创新是“新视角预测”，这使其与大型语言模型的“下一个词元预测”以及视频模型的“下一帧预测”区别开来。 Atlas 接收给定数量的场景视图或描述，并创建一个“空间上下文”。从这个上下文出发，它能够预测在空间和时间中任何任意虚拟摄像机位置下，世界应该呈现出怎样的景象。这使得“摄像机条件生成”等功能成为可能，用户可以输入一张图像和一条摄像机轨迹，从而从任何期望的视角生成视频帧。它在“稀疏三维重建”方面也表现出色，能够从少至一帧、多达一百帧的数据中重建真实世界，然后将其输出为新颖的视频或显式的三维模型。在模拟方面，Atlas 可以创建类似于《黑客帝国》中的“子弹时间视频”和机器人模拟。 Atlas 在稀疏三维重建方面取得的一个重大突破是，将复杂镜头所需的摄像机数量从数百个（如《黑客帝国》原版“子弹时间”场景中）减少到仅需三个。这消除了工作室拍摄、绿幕或昂贵校准的需要，与传统的“密集重建”相比，数据输入量“减少了 50-100 倍”。NeRF 的创建者 Ben 解释说，密集重建传统上需要从多个角度捕捉每一个点，这导致了繁琐且耗时数小时的工作，以避免三维模型中出现空洞。相比之下，Atlas 利用生成能力“想象”并填补数据稀疏处的空白，使得三维重建变得更加易于实现和高效。 Feifei 强调，Atlas 首次实现了在单一模型中“像素生成和像素重建的统一”，这个问题在历史上一直由计算机视觉领域内不同的子领域分别解决。Atlas 是原生多模态的，从一开始就设计为能够将文本、图像、视频、摄像机姿态和三维数据（如深度图）作为集成输入进行处理。团队对“规模化假设”抱有坚定的信念，指出扩大模型规模并延长训练时间，其性能始终得到提升，这表明他们仍处于其潜力的“初期”，目前仅受限于计算资源。 在应用方面，Atlas 拓展了 World Labs 此前的产品 Marble，提供了一种更强大、更统一的方法。对于创意人士而言，它为生成内容提供了三维一致的基准，实现了对故事板制作、电影制作和游戏环境设计的精确控制。对于设计师和建筑师而言，它简化了虚拟设计阶段，通过将口头或基于草图的想法转化为具体的3D模型，从而实现更快的迭代和反馈。 机器人技术是一个关键的应用领域。Feifei 解释说，Atlas 是“现实到模拟”管线的关键。机器人训练是数据密集型的，通常需要大量的现实世界数据和模拟环境中的随机化。Atlas 的稀疏重建能力极大地加速了从现实世界扫描数据创建模拟环境的过程，取代了“极其痛苦”的密集重建方法。这有助于克服机器人技术中的数据瓶颈，从而开发能够理解世界如何对动作做出反应的“神经模拟器”，进而改进机器人策略训练。 展望未来，团队承认动态是关键前沿。Atlas 已经融入了从其训练数据中学习到的“初级动态”，使其输出中包含移动的汽车或水波等元素，这是静态的 Marble 模型无法实现的。未来的开发将侧重于增强动态能力，探索“4D 视频”（3D + 时间），并扩展“多模态控制”和可编辑性。最终目标是允许用户与生成场景的各个方面进行交互和控制，从布局到物体身份和时间。发言人坚信“新视角预测”是“AI 完备”的，类似于下一个词元预测，这与大自然赋予动物眼睛以体验新视角的原理不谋而合。

World Labs has recently launched Atlas, a new "frontier model" that has received an "amazing reception." Feifei, Justin, and Ben discussed its capabilities, significance, and future implications. Atlas is described as a "next-generation world model" that can generate, reconstruct, and simulate the world. Its core innovation is "new view prediction," differentiating it from LLMs' "next token prediction" and video models' "next frame prediction." Atlas takes a given number of views of a scene or a description and creates a "spatial context." From this context, it can predict what the world should look like from any arbitrary virtual camera position in space and time. This enables capabilities such as "camera condition generation," where users can input an image and a camera trajectory to generate video frames from any desired perspective. It also excels at "sparse 3D reconstruction," allowing the reconstruction of the real world from as few as one to up to a hundred frames, which can then be output as a novel video or an explicit 3D model. For simulation, Atlas can create "bullet time videos," reminiscent of the Matrix, and robotic simulations. A significant breakthrough Atlas achieves in sparse 3D reconstruction is reducing the number of cameras needed for a complex shot from hundreds (as in the original Matrix "bullet time" scene) to just three. This eliminates the need for studio capture, green screens, or expensive calibration, representing a "50-100x reduction" in data input compared to traditional "dense reconstruction." Ben, creator of NeRF, explained that dense reconstruction traditionally required capturing every single point from multiple angles, leading to tedious, hours-long efforts to avoid holes in the 3D model. Atlas, by contrast, uses generation to "imagine" and fill in gaps where data is sparse, making 3D reconstruction much more accessible and efficient. Feifei highlighted that Atlas represents the first time there's been a "unification of pixel generation and pixel reconstruction" in a single model, a problem that historically has been tackled by separate subfields in computer vision. Atlas is natively multimodal, designed from the start to work with text, images, videos, camera poses, and 3D data (like depth maps) as integrated inputs. The team expressed strong conviction in the scaling hypothesis, noting that making the model bigger and training it longer consistently improved performance, suggesting they are "at the beginning" of its potential, currently limited only by compute resources. Regarding applications, Atlas extends World Labs' previous product, Marble, by offering a more robust and unified approach. For creatives, it provides a 3D-consistent grounding for generations, enabling precise control for storyboarding, movie production, and game environment design. For designers and architects, it streamlines the virtual design phase, allowing quicker iterations and feedback by translating verbal or sketch-based ideas into concrete 3D models. A crucial application area is robotics. Feifei explained that Atlas is key to the "real-to-sim" pipeline. Robotics training is data-intensive, often requiring vast amounts of real-world data and randomization in simulated environments. Atlas's sparse reconstruction capabilities dramatically accelerate the process of creating simulated environments from real-world scans, replacing the "excruciatingly painful" dense reconstruction methods. This helps overcome the data bottleneck in robotics, allowing for the development of "neural simulators" that can understand how the world responds to actions, thereby improving robotic policy training. Looking ahead, the team acknowledged that dynamics are a key frontier. Atlas already incorporates "baby dynamics," learned from its training data, allowing for elements like moving cars or water waves in its outputs, which was not possible with the static Marble model. Further development will focus on enhancing dynamic capabilities, exploring "4D video" (3D + time), and expanding "multimodal control" and editability. The ultimate goal is to allow users to interact with and control various aspects of the generated scenes, from layout to object identity and time. The speakers strongly believe that "new view prediction" is "AI complete," analogous to next token prediction, fundamentally aligning with how nature gave animals eyes to experience new viewpoints.