World Labs Launches Atlas for 1440p Camera-Controlled Video, 3D Worlds and Robot Views

World Labs has launched Atlas, a model that promises to turn a few reference images into controlled 1440p video, explicit 3D scenes and simulated views from a robot’s cameras. The pitch is unusually broad; the immediate evidence is narrower, resting on company-reported tests while access begins with selected partners rather than a public API.

The camera path becomes an input

Atlas is a multimodal autoregressive diffusion transformer, a model design that generates outputs step by step while using diffusion to produce images and video. It accepts text, images, camera poses and 3D depth maps, placing those inputs into a shared spatial context: each image is associated with a location in three-dimensional space.

That architecture makes camera geometry a native control rather than an instruction interpreted from words such as pan or crane. World Labs says users can supply one to six reference images and a manually designed path to generate as much as one minute of 1440p video. The model also fills in parts of a scene outside the supplied views, which gives it creative range but makes the boundary between reconstruction and invention important.

75%Camera-control preference

World Labs reported that human raters preferred Atlas in 75% to 94% of camera-controlled generation trials, depending on the competing model.

25.3Sparse-view reconstruction error

World Labs reported a lower error score for Atlas than the next-best compared open-source baseline; lower is better on this measure.

More images constrain the model’s imagination

Atlas can reconstruct a real location from one or more images, generate views from unseen angles, and export point clouds or 3D Gaussian splats. Point clouds represent a scene with spatial points; Gaussian splats are a 3D representation used to render those scenes. World Labs says two or three images will typically produce a faithful reconstruction, while more images reduce how much scenery it must infer.

That trade-off is central to the product’s practical value. A model that convincingly invents a missing side of a building can help create a film shot or game environment. It is less useful when a team needs a reliable replica of a factory or workspace. Atlas’s stated answer is to add more source material, but its public examples do not settle how reliably it handles inconsistent images or unfamiliar environments.

Two routes from spatial context to a product

• For visual effects, World Labs says Atlas can reframe footage from as few as three cameras, creating new angles after a moment has been captured.
• For robotics, it says Atlas can reconstruct an environment and generate the RGB images and depth readings a simulated robot-mounted camera would observe along a route.
• For manipulation simulation, World Labs says the system supports scenes involving rigid, articulated and deformable objects, alongside changes to objects, robot motion, lighting and backgrounds.

Partner workloads now carry the burden of proof

World Labs’ benchmark claims are promising but not yet independently established. Its internal evaluations found stronger camera-motion adherence than competing models, and its sparse-view test put Atlas at 25.3 mean absolute-relative pointmap error against 28.7 for the next-best compared open-source baseline. The reconstruction comparison did not include closed commercial systems.

Atlas is entering early access through a request form, with selected partners first in line. World Labs has not disclosed partner identities, pricing or a general-availability date, and says Atlas will power future versions of Marble and other products. The launch therefore establishes the model’s design and ambitions; partner use will show whether one spatial context can hold up across the much messier demands of production footage and robot interaction.