World Labs releases Atlas tools for editable 3D scenes World Labs releases Atlas tools for editable 3D scenes

A route from images to usable scenes

The release includes Viewfinder for building scenes from images and Chisel for controlling their structure. Splats to Mesh converts the representation into geometry, while Relight changes illumination. The developer API connects these operations to other applications. For a team making a virtual set, the important change is the handoff: a generated view can be part of a scene that moves into an existing production workflow. A finished image, an editable environment and an exportable object solve different parts of that production job, with different deliverables.

The model predates this release

Atlas itself was introduced on September 1. World Labs described a model that combines text, images, camera positions and depth in one spatial context. Instead of asking for a vague pan in a prompt, its camera input specifies where the next view should come from. The company demonstrated a one-minute video at 1440p with a planned path. The November release packages capabilities around that earlier research. A camera path also changes what consistency means: the same doorway must remain in the same place as the viewpoint moves, rather than merely looking convincing in an isolated frame.

Reconstruction and invention share a pipeline

World Labs’ September examples distinguish visible evidence from invented surroundings. One photograph of a garden produces a wider scene containing imagined buildings. Additional photographs supply context for the cottage and main house. The model can also output point clouds and Gaussian splats, a representation used to render a scene from different directions. This creates two legitimate uses with different standards. A fictional set benefits from plausible additions. A representation of an existing location needs those additions separated from captured detail. A smooth camera move alone does not establish that an unseen wall, passage or surface exists in the photographed place.

Interactive exploration is a separate tradeoff

On November 5, World Labs introduced Frameboy, a research preview streaming generated views from cloud GPUs at up to fifteen frames per second. It also described a Viewfinder overlay for exploring user-built scenes. At that point Viewfinder was in closed beta. The preview focused on changing viewpoints quickly instead of waiting for each manually planned sequence. World Labs noted that this mode required a powerful cloud GPU server and cost more than local rendering methods. The distinction remains useful when choosing a workflow: exploring a scene, producing a final camera sequence and exporting geometry are separate operations. Faster feedback helps select a view; the exported asset still has to suit the project that receives it.