Skip to content
vrBlogs

World Labs Atlas API Brings Splats to Mesh Workflows

World Labs’ Atlas API exposes six task endpoints for scene generation, splat reconstruction and GLB mesh conversion, with documented inputs and pricing.

v

By vrBlogs

· 4 min read · Source: VR.org News

World Labs’ Atlas API provides six task endpoints for scene generation, Gaussian-splat reconstruction and conversion of splats into GLB meshes. The API reference lists the endpoints, request paths and supporting asset and operation services.

VR.org News reported that developers can use Atlas through a project API key following World Labs’ November 12 release post. Atlas was introduced on September 1. Atlas is World Labs’ world model: it takes photographs or a text prompt and builds a scene that remains consistent in 3D, allowing it to render new views, a video or a splat that users can walk through.

Four apps built on Atlas

Four apps are available to anyone who signs in. Viewfinder reconstructs a scene from photographs or generated images and lets users fly a camera path through it. Chisel lets users graybox a scene and see every camera view update while they move the blocks.

Relight relights and reworks a splat scene. Splats to Mesh converts a splat into a mesh, as its name indicates.

Six task endpoints and asynchronous jobs

The six task endpoints are atlasChisel, atlasGenerate, atlasTextToImage, images2PosedRGBD, posedImages2Splats and splats2Mesh. Each uses a POST request under /api/v2/tasks, followed by the task name, according to the Atlas API reference.

Developers create a project key, configure a client or sign in from the World Labs CLI, then submit a task. Atlas returns an operation instead of the finished result. Developers can poll /api/v2/operations/{operation}, call :wait, cancel work with :cancel or request :trace, which returns the exact submitted request beside its result.

Asset endpoints handle file uploads and outputs. The API also provides a health check, an assets collection for uploading and reading files, an operations collection, and webhooks that can be set on tasks.

Why splats2Mesh matters for VR

splats2Mesh converts Gaussian splats into what World Labs calls a mesh “with real geometry,” returned as a GLB. A Gaussian splat stores a scene as millions of soft, coloured ellipsoids rather than triangles. The format renders captured rooms quickly on a desktop GPU, but splats have no surface for ordinary collision, and engine lights and shadows do not affect them like they affect triangle meshes.

The task can fit upright collider boxes around the mesh and return them as a second GLB. The result can be passed to Unity or Godot, where collision can work with mesh geometry. A three-call workflow can start with images2PosedRGBD to estimate cameras and depth, continue with posedImages2Splats, then use splats2Mesh to produce geometry. The splat step usually takes 15 to 25 minutes.

Splat-to-mesh conversion predates Atlas. SuGaR did it in late 2023, and KIRI Engine added a one-tap version to its phone app in March 2024. The reported difference is access from custom code without separately wiring together COLMAP, a splat trainer and a meshing step.

Tools, industries and engine integration

Blender is the only named 3D tool on the Atlas page. The documentation also lists an MCP server for Claude and other agent clients, plus Spark.js for rendering splats on the web. The report names no Unity package, Unreal plugin or Godot addon.

Atlas lists film, architecture and robotics as industries, while pitching Splats to Mesh for games and interactive environments. Its returned GLB is a format Godot imports natively, and Unity reads it through its glTFast package. Godot users already lean toward that format; the Blendkit plugin added a prefer-GLTF option for the same reason. No editor plugin has yet been written to call the six endpoints from inside an engine.

Inputs, outputs and documented costs

The posedImages2Splats documentation says refined output costs 5,000 credits, or $5. Draft output with refine: false costs 500 credits, or $0.50. Refined runs accept 2 to 512 ordered posedFrames; drafts accept 2 to 32.

Every frame combines an image asset, camera and depth. The depth must be a linear-depth EXR under depth.depthAsset, as returned by images2PosedRGBD or by atlasGenerate with returnDepth. An asset can be a project-owned asset ID, inline base64 smaller than 10 MiB or a public HTTPS URL. Each returned image, camera and depth must remain together.

Every camera grid must use one of 1280x720, 720x1280, 960x720 or 528x704, measured as width by height, and all frames must declare the same grid. RGB images are resampled to that grid, so camera intrinsics describe the selected grid rather than necessarily the original image. Outputs from images2PosedRGBD and atlasGenerate at those four grids can pass directly into the workflow. The result of posedImages2Splats is an SPZ file.

Atlas billing documentation sets the conversion at 1,000 credits per US dollar. Credits are drawn from the account balance belonging to the project’s API key. Account owners and billing owners can add credits manually or enable automatic recharge; payment methods and invoices are managed under Settings → Billing, while balances, top-ups and monthly spend limits are under Settings → Developer Platform.

Other task rates appear in the signed-in platform and may change, so batch estimates should use its current rate card. The report says there is no public full Atlas pricing page. It also says World Labs signed a definitive agreement on September 28 to be acquired in an all-stock deal valued; the supplied report text ends at that word before stating the valuation.

Filed underVR AI
Share

Enjoyed this story?

Like it to tell the editors, or save it to read later.

Comments

Sign in to join the conversation.

Sign in

    No comments yet. Start the conversation.

    All news

    Are you sure?