World Labs Debuts Atlas, a Multimodal AI Model for Spatial Intelligence

World Labs has introduced Atlas, a next-generation world model designed to work natively across text, images, video and 3D. The model can generate controlled video, reconstruct physical environments, simulate scenes over time and produce explicit 3D outputs, with early access now opening to select partners.

Atlas is built as what World Labs calls an omni model, using a multimodal autoregressive diffusion transformer architecture. Instead of treating images, video and 3D data as separate inputs, the system combines them into a shared spatial context and generates new outputs based on that context.

One of Atlas’ main capabilities is camera-controlled generation. Users can provide between one and six reference images, define a camera path and generate new views that remain consistent with the geometry of the source material. World Labs says Atlas can produce videos up to one minute long at 1440p resolution.

The model accepts camera geometry directly rather than relying only on text descriptions of camera movement. That allows users to specify exact viewpoints and motions while Atlas fills in parts of a scene that were not visible in the original images.

Atlas can also combine unrelated reference images by placing them within the same 3D context. The model then generates transitions and connective spaces between those inputs, effectively building a larger environment around the supplied material.

For reconstruction tasks, Atlas can work from as little as a single image or use more than 100 images when greater fidelity is needed. World Labs says the model can generate both novel 2D views and explicit 3D representations of real spaces. Those 3D outputs include point clouds and Gaussian splats. Atlas can estimate geometry from still images or video, fill in unseen areas and turn the result into a 3D scene that can be rendered interactively.

World Labs says the same capabilities can be applied to robotics through Real-to-Sim workflows. A physical environment can be captured with ordinary video, reconstructed in 3D and then used as the basis for robot simulation. As a simulated robot moves through the reconstructed environment, Atlas can generate the RGB images and depth data that onboard sensors would be expected to observe. The system can also be used to simulate object manipulation and vary elements such as object placement, lighting, robot motion and backgrounds.

Atlas also supports video reframing. World Labs says footage captured from as few as three cameras can be reconstructed so that users can freeze an event and view it from different angles without a specialized capture setup.

Image generation is another part of the model, though World Labs describes it as secondary to the broader world-modeling focus. Atlas can generate images and 360-degree panoramas from text or image prompts, follow complex instructions and render text across different visual styles.

The company says Atlas outperformed several specialized models in its internal evaluations of camera-controlled generation and 3D reconstruction. In third-party human testing focused on following camera paths, Atlas was preferred over MiniMax H3, Gemini Omni Flash, Happy Horse 1.1, FLUX 3 and Seedance 2.5.

World Labs also reported lower average reconstruction error than a group of open-source 3D reconstruction models across several benchmarks. The company said those results were produced under a common evaluation protocol.

Atlas was pretrained from scratch on a large multimodal dataset. World Labs says its performance improved as training compute increased and expects further gains as the model continues to scale.

The model will also serve as the foundation for future versions of Marble and other World Labs products. For now, Atlas is entering early access with selected partners, and the company has not announced a date for broader availability.

See more here:

This analysis is based on reporting from World Labs.

Image courtesy of World Labs.

This article was generated with AI assistance and reviewed for accuracy and quality.