World Labs presents Atlas: the new world model for spatial intelligence

On September 1, 2026, World Labs announced Atlas, a new world model designed to understand, reconstruct, and simulate three-dimensional environments.

Its goal goes beyond simply generating images or videos: Atlas aims to build a coherent representation of space, allowing artificial intelligence to generate new viewpoints, reconstruct real-world scenes, and simulate environments over time.

What is Atlas?

Atlas is a multimodal model pre-trained to work natively with text, images, videos, and 3D data.

All of this information is integrated into a single shared spatial context.

Unlike traditional generative models, images are not treated as independent frames: they are positioned within three-dimensional space, allowing the model to maintain geometric consistency and imagine parts of a scene that were not directly observed.

Atlas can also generate explicit 3D geometry, such as point clouds and 3D Gaussian Splats, in addition to images and videos.

Key capabilities

One of Atlas's most interesting features is precise camera control.

Starting from one or more images, Atlas can generate new views and videos while maintaining the three-dimensional consistency of the scene.

The model can also reconstruct real-world environments from just a few images. When the available information is limited, it fills in the missing parts using its knowledge of the world; as more images become available, the reconstruction becomes progressively more accurate.

Atlas also works across the temporal dimension.

It can reinterpret a scene from new perspectives, create bullet-time-like effects, and support Real-to-Sim simulation applications for robotics, generating virtual environments from observations of the real world.

Despite its focus on 3D, Atlas can also generate images from text, 360-degree panoramas, and content in a variety of styles.

For more information, visit the official page here.