World Labs announced Atlas, a new model that the company describes as a “world model” for spatial intelligence. Atlas combines text, images, video and 3D data within a single architecture, with the aim of understanding spatial relationships between elements rather than merely generating separate images or video clips.
Fei-Fei Li, an artificial intelligence researcher, co-founded World Labs, which was established in 2024 to focus on enabling AI systems to understand the physical world and 3D relationships. The company says Atlas was pretrained from scratch and technically relies on an architecture that combines an autoregressive multimodal diffusion model and a Transformer.
What Does Atlas Offer?
The model can combine different types of inputs within a shared spatial context, allowing it to estimate unseen parts of a scene based on the 3D structure of objects and environments. One of its notable features is treating the camera’s position and movement path as direct inputs, rather than relying solely on describing camera movement through text prompts.
Using one to six reference images and predefined camera paths, Atlas can generate video at 1440p resolution lasting up to one minute. It can also estimate the shape of an object or environment from different angles based on a single image, and expand a scene to include parts not visible in the original image. The tool also allows unrelated images to be placed within their positions in 3D space and transitions to be created between them within a single environment.
Reconstructing the Real World
World Labs says Atlas can be used to reconstruct real spaces in 3D. Two or three images may be sufficient to recreate many scenes, while providing dozens or more than one hundred images reduces the need to guess unseen areas and brings the result closer to the original environment.
The outputs are not limited to images and video; the model can also produce 3D point clouds and files in the “3D Gaussian splat” format, outputs that can be used in games, design, visual effects and robotics.
Robotic Simulation and Current Limitations
Phone recordings of a real environment can also be used to create a 3D simulation, after which the data that virtual robot cameras and depth sensors might see during movement can be generated. The simulation allows the positions of objects, lighting, background and robot movement to be changed to create multiple scenarios for training and testing.
Why does this announcement matter? Atlas connects content generation with spatial understanding and simulation, making it closer to a tool for building explorable environments that can be used in robotic systems, rather than merely a video-generation model. World Labs says the model will form the foundation of the Marble platform for creating 3D worlds and other company products.
However, Atlas remains in an early-access phase for selected partners. The company has not disclosed its price, public availability date, model size or the computing capacity used to train it. The claim that it outperforms some specialized open-source models in 3D reconstruction from scattered images is also based on the company’s tests and has not yet been independently verified by researchers.
In February 2026, World Labs raised $1 billion in funding with participation from AMD, Autodesk, Nvidia, Fidelity and other investors, bringing its total funding to approximately $1.23 billion.