Atlas by World Labs builds explorable 3D scenes from images: you pose one or more photos, generate new camera views around them, then turn those views into a video or a Gaussian splat. The web app at atlas.worldlabs.ai wraps this in tools such as Viewfinder and Chisel, and the same steps are available through a developer API. This guide walks through a first project and covers the mistakes that most often waste credits, as of November 2026.
What Atlas is (and what it is not)
World Labs describes Atlas as its "next generation world model" for generating and reconstructing 3D worlds. The launch post says it works natively with text, images, video and 3D, can output up to 1 minute of camera-controlled video at 1440p, and can reconstruct a scene from as few as two or three images.
Atlas is scene-level. It is built for rooms, streets, landscapes and architecture, not for single props. World Labs' best-practices page says the models are not yet as strong on people or moving objects, so treat crowds, close-up faces and moving vehicles as weak spots.
The web app lists these tools at launch:
| Tool | What it does (per World Labs) | Our guide |
|---|---|---|
| Viewfinder | Reconstructs a scene from photos or generated images, explores new viewpoints, renders camera paths | Viewfinder guide |
| Chisel | Builds and grayboxes a scene so you control its structure | Chisel guide |
| Relight | Explores new ways to light and manipulate splats | Relight guide |
| Splats to Mesh | Converts splats into meshes for games and interactive environments | Splat to mesh guide |
| Developer API | Connects Atlas to your own apps and tools, including Blender | Atlas API guide |
Step 1: Create an account and check your credits
Sign in at atlas.worldlabs.ai. Billing works in credits. For the Developer API the documented rate is 1,000 credits per US dollar. Balances, top-ups and monthly spend limits are under Settings, and the Usage page shows what each day cost, split into Studio and Developer usage.
Free credits: a new account we checked in November 2026 received 10,000 promotional Studio credits and 10,000 "Developer Signup Credits" for the API (about $10). Both grants showed an expiry one year after signup. The Studio page says plan credits expire at the end of the billing cycle and top-ups expire one year after purchase. Grant amounts can change, so check your balance after signing in.
Some Studio actions have listed prices:
| Studio action | Credits |
|---|---|
| Prompt enhancement | 20 per request |
| Image edit | 125 per request |
| Image upscale | 10 per image |
| Chisel caption | 20 per request |
| Frameboy session (up to a minute) | 500 per session |
| Relight image edit | Whatever the model provider charged, converted to credits |
API tasks are priced in the Atlas API pricing table in the World Labs console (checked November 14, 2026). Frame prices are for standard 720p output, and RGBD includes the depth surcharge:
| API task | Credits |
|---|---|
| Images to posed RGBD | 2 per input image ($0.002) |
| Atlas image generation | 40 RGB / 50 RGBD per output image |
| Atlas view generation | First frame 40 RGB / 50 RGBD, each extra frame 8 / 10 |
| Atlas Chisel | First frame 40 RGB / 50 RGBD, each extra frame 8 / 10 |
| Splats, refined / draft | 5,000 ($5) / 500 ($0.50) per request |
| Mesh export | 500 per world ($0.50) |
| Third-party model calls | Passed through at the provider's public rate, converted |
Our pricing calculator estimates a project from these numbers.
Step 2: Pose your images in Viewfinder
Every Atlas project starts with posed images. Posing estimates where each photo was taken, which way it faces, its field of view, and a depth map. There are two ways to start:
- Reconstruct a real place. Import several overlapping photos of the same place together. In Viewfinder the docs call this action Reconstruct. Use sharp photos of a mostly static scene with similar exposure.
- Start from one image. A single photo or a generated image is enough. Atlas invents what lies outside it.
If you are working from video, export sharp still frames first. Atlas takes images, not video uploads. For a single room, World Labs suggests 8 to 16 well-distributed frames with roughly 50% overlap between neighbors.
Step 3: Generate views around the scene
Next, generate a small ring of views around the part of the scene you care about. The workflow guide suggests 6 to 12 new cameras in one request:
- Spread the cameras around the main subject, each looking at it, at about the height of the original camera.
- Keep every camera in space the photo shows as empty. A camera inside a table or behind a wall produces large invented shapes.
- Give each camera its own position. Cameras that only rotate in place cannot recover depth.
- Generate all the views together, so Atlas decides the unseen parts once.
From a single image, a camera can turn about half the field of view, or move about a third of the typical depth, and still show mostly what the image showed. Further than that, Atlas is inventing. In World Labs' potter's-studio example, seven views around a work table took 23 seconds.
Step 4: Check the views before the expensive step
This is where you save credits. Look at every view. Objects should keep their shape, position and appearance from one view to the next. If a table is round in one view and square in another, remove the view that disagrees, or generate a replacement with the good views as context.
The splat and video steps build on these views and cost more, so problems you miss here carry through.
Step 5: Render a video along a camera path
To make a flythrough, author a camera path in Viewfinder, generate it, and export the video. The docs give these rules for a clean path:
- Start on a camera the viewer knows. To open on your photo, make the first camera its posed camera.
- Keep the path in open space and near where your views were generated. Even a straight line between two open positions can pass through a door frame.
- Use one camera per frame (seconds times frame rate), evenly and closely spaced.
- Render the whole path in one request. Frames from separate requests do not match.
World Labs' example used 192 cameras for 8 seconds at 24 frames per second and finished in 109 seconds. Through the API, Atlas returns image frames rather than a video file, and the docs assemble them into an MP4 with ffmpeg.
Step 6: Build a Gaussian splat
For a 3D scene you can walk through, reconstruct the checked views as a Gaussian splat. The API task is posedImages2Splats:
| Mode | Cost | Frames accepted |
|---|---|---|
| Refined (default) | 5,000 credits ($5) | 2 to 512 |
Draft (refine: false) | 500 credits ($0.50) | 2 to 32 |
Failed runs are free. Timing is reported differently across pages: the task guide says a run usually takes 15 to 25 minutes, while the API reference says 10 to 20 minutes. The worked example took 18 minutes and returned 2,000,000 splats. Either way, submit once and wait. Submitting again while a run is going starts a second run and a second charge.
If you would rather skip the setup, our own photos to 3D app runs a photos → splat → GLB pipeline and is paid with this site's credits.
Step 7: Export to Blender, the web or a game engine
The splat output is an SPZ file in the same world space as your cameras. From there:
- Blender 5.3 imports
.spzor.plysplats natively. Drag the file into the viewport. - Web: render the SPZ with Spark in three.js or React.
- Game engines: convert the splat to a textured GLB with Splats to Mesh, then import it into Unity or Unreal. See Splat to mesh.
Tips to avoid failed renders
- Move the camera, don't just turn it. World Labs recommends capturing views with translation, not only rotation. Through the API, a request whose cameras all share one center is rejected when you ask for depth.
- Keep cameras out of solid objects. Check depth first to find the floor and open space.
- Prune disagreeing views. Doubled, ghosted objects in a splat almost always come from two views that disagree.
- Cover what you will look at. Floors, low furniture and the far side of objects need at least one view, or they come out blurry or invented.
- Expect weak spots. Sky, foliage, glass, thin structures like chair legs, and faces are listed as hard cases.
- Watch spend limits. When a monthly limit reaches 100%, new Developer API tasks pause with a
402error.
Where to go next
- Learn the tools in depth: Viewfinder, Chisel, Relight, Splat to mesh.
- Building an app? Read the Atlas API guide.
- Not sure whether you need Atlas or Marble? See Atlas vs Marble.