Atlas Viewfinder Guide: Image to 3D Scene

Use Atlas Viewfinder to turn one image into an explorable scene, render new camera angles, create Gaussian splats and export MP4, PLY or SPZ.

Updated 2026-11-14 · Independent guide, not affiliated with World Labs

Viewfinder is the Atlas tool that turns a photo, or several photos of one place, into a scene you can move through. It estimates a camera and depth for each image, generates new viewpoints around it, renders camera paths as video, and can turn the result into a Gaussian splat. Here is how to use it well and why renders fail, based on World Labs' documentation as of November 2026.

What Viewfinder does

World Labs' Atlas page describes Viewfinder in three parts: it reconstructs a scene from photos or generated images, lets you explore new viewpoints, and renders camera paths. Radiance Fields reported that Viewfinder is the renamed Atlas Preview panel from Marble Studio.

Under the hood, Viewfinder follows the same steps the docs describe for the Developer API:

StepDeveloper API taskWhat you get
Pose the imagesimages2PosedRGBDA camera and depth map for each image
Explore new viewsatlasGenerateNew RGB views, optionally with depth
Render a pathatlasGenerate with an ordered camera pathOne frame per camera, assembled into a video
Create splatsposedImages2SplatsAn SPZ Gaussian splat

Knowing this mapping helps when something goes wrong, because the API docs spell out the rules the app has to follow.

Step 1: Choose an image that suits Atlas

Atlas does best on spaces: rooms, architecture, landscapes, the relationships between surfaces. World Labs says the models are not yet as strong on humans or dynamic objects. A good input:

  • is sharp, with a clear environment and plenty of visible surfaces;
  • is a wide or medium view, not an extreme close-up;
  • shows a mostly static scene with consistent exposure.

Crowds, close-up people, moving vehicles and moving water reduce consistency. People can appear in the shot, but they should not be what you care about most.

Step 2: Import and reconstruct

Start a new Viewfinder project and import your image. With one image, Atlas estimates its camera and depth, and later invents what lies outside the frame.

With several photos of the same real place, import them together and use Reconstruct. The docs name that action for frames pulled from a video walkthrough. Reconstructing them together places all the cameras in one shared world. For a single room, World Labs suggests starting with 8 to 16 well-distributed images that share roughly 50% of their content with their neighbors.

If your source is video, export sharp still frames first. Atlas takes images, and direct video upload is not available.

Step 3: Place cameras and generate new views

This is the creative part. You choose where Atlas should look next, and it renders those viewpoints. The workflow guide suggests:

  • 6 to 12 new cameras in one request;
  • cameras spread around the main subject at about the original camera's height, each looking at it;
  • every camera in space your image shows as empty, never inside an object or behind a wall;
  • each camera at its own position, with neighboring views overlapping.

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 mostly show what the image showed. Beyond that Atlas invents. In World Labs' example, seven views around a potter's work table took 23 seconds.

Step 4: Inspect before you continue

Look at every generated view. Objects should keep the same position, shape and appearance from view to view. Delete a view that disagrees, or generate a replacement with the good views as context. These views are the input to both the video and the splat, so they work as a preview of the final result.

Step 5: Render a camera path and export video

For a flythrough, the docs say to author a camera path in Viewfinder, generate, and export the video. A finished splat is not needed for this step.

Path rules from the docs:

  • Open on a camera the viewer recognizes, for example your photo's own camera.
  • Stay in open space and close to where the views were generated.
  • Use one camera per frame, evenly spaced (seconds times frame rate).
  • Render the whole path at once. Frames from separate requests don't match and can flicker.

In World Labs' example, 192 cameras (8 seconds at 24 fps) rendered in 109 seconds. The launch post says Atlas can output up to 1 minute of video at 1440p. Viewfinder's card in the web app lists MP4 as its video format; the docs do not state the render settings. Through the API you get image frames and assemble an MP4 yourself, for example with ffmpeg.

Step 6: Create splats and export SPZ or PLY

To get a navigable 3D asset, turn your checked views into a Gaussian splat. Through the API this is posedImages2Splats: a refined run costs 5,000 credits ($5), a draft costs 500 credits ($0.50), and failed runs are free. Radiance Fields reported the same 5,000-credit price for Viewfinder's Create splats option, with a run of about 20 minutes. The Viewfinder card in the web app confirms PLY and SPZ as splat outputs.

The docs disagree slightly on timing: 15 to 25 minutes in the task guide, 10 to 20 minutes in the API reference. Their worked example took 18 minutes and returned 2,000,000 splats.

What to do with the file:

  • Open it in Blender 5.3, which imports .spz and .ply splats natively.
  • Render it on the web with Spark.
  • Convert it to a GLB mesh with Splats to Mesh for Unity or Unreal.

If you want a hosted shortcut from photos to a splat and a GLB, our photos to 3D app runs that pipeline using this site's credits.

What Viewfinder costs

View generation is priced per output frame in the Atlas API pricing table in the World Labs console (checked November 14, 2026): 40 credits for the first frame and 8 for each additional frame in RGB, or 50 and 10 in RGBD (the depth surcharge is included). Prices are for standard 720p output, at 1,000 credits per $1. Posing your photos costs 2 credits ($0.002) per input image. A 192-frame RGBD path works out to 50 + 191 x 10 = 1,960 credits (about $1.96). Use our pricing calculator for an estimate.

Common failure reasons

Most failed or bad Viewfinder results come from a short list of causes.

Rotating in place instead of moving

Atlas recovers depth scale from how far apart the cameras are. The API rejects a request whose cameras all share one center (a "pure-rotation rig") with HTTP 422 when depth is requested. A one-view request with depth output is also rejected unless you set depth off or add at least two translated targets. In the app, the fix is the same: move the camera, don't only turn it.

Cameras inside objects or behind walls

A camera placed inside furniture or behind a wall produces large invented shapes in front of the lens. Use the depth map to find the floor and open space before placing cameras.

Paths that clip through geometry

A path between two open positions can still pass through a door frame. In World Labs' own example, the first version of a path looked into the door frame. Moving straight into the room first fixed it.

Views that disagree

If a table has a different shape or position in two views, the splat combines them into a doubled, ghosted object. World Labs shows a cabin where this happened to a dining table.

Gaps in coverage

Areas no view shows come out blurry or invented. A common miss is the floor: in one lobby example the views looked mostly at the ceiling, and the lower part of the scene filled with blurry furniture.

Hard materials and subjects

Sky, foliage, large plain walls, windows, thin structures like chair legs, and faces can all reconstruct poorly even with good inputs. Check these from the angles you plan to use.

Inconsistent source photos

Blurred frames, changing exposure, or people who moved between shots break reconstruction. Remove those frames rather than adding more.

Next steps

  • Shape a scene from a layout instead of a photo with Chisel.
  • Take your splat into a game engine with Splat to mesh.
  • See the full workflow in the Atlas guide.

FAQ

Related

Sources