World Labs Atlas: a world model that builds scenes in 3D
Atlas is World Labs’ new world model, introduced on September 1, 2026. It brings text, images, video and camera data into one shared 3D context, then uses it to create new viewpoints, long camera moves and reconstructions. Its key idea is not just attractive motion: the scene is meant to remain spatially consistent when the camera travels beyond the original frame.
What Atlas can do
- Generate images and videos from one or more references with a specified camera path.
- Reconstruct a real space from sparse photos or video and show new views.
- Produce explicit 3D outputs, including point clouds and Gaussian splats.
- Reframe events filmed by several cameras and create simulation scenes for robotics.

Why it differs from ordinary image-to-video
Each image is grounded at a position in space, so creators can specify camera positions and angles instead of relying only on a written instruction. Visible details anchor the result; unseen areas are plausibly completed. More reference images give the model more evidence and less room to invent.
Video: from one frame to a spatial scene
This original World Labs demonstration expands a reference into new views and a 3D representation. The clip is served from our domain; the source is linked below.
World Labs Atlas demonstration: new views and a 3D scene from one image
Reconstruction, VFX and robotics
Atlas can work from one to dozens of photos. World Labs says two or three images often establish a useful reconstruction, while larger sets improve fidelity. For VFX, a few ordinary cameras can be reframed into an impossible viewpoint. For robotics, the model connects a spatial reconstruction to the RGB and depth observations a robot would receive along a route.
Who should watch Atlas
3D artists, previs directors, VFX teams and robotics researchers all benefit from a reusable world rather than a single generated frame. Atlas is in early access for select partners; access can be requested from World Labs.
FAQ
When was Atlas released?
World Labs introduced Atlas on September 1, 2026.
What inputs does it use?
Text, images, video as image sequences, camera poses and depth maps.
Where are the original visuals?
The images and video used here come from the World Labs Atlas announcement.