← Applied Spatial Intelligence Labs

South House — a building, rebuilt in 3D from photographs

No scanner, no LiDAR, no drone. Just 64 ordinary photos walked around the structure. Our spatial pipeline recovers where every camera stood and optimizes the scene into a photoreal 3D Gaussian-splat model — a dense field of oriented, colored splats rendered live in your browser. The same reconstruction core behind ASI Labs digital twins.

Model · gaussians Input · photos Runs · in-browser WebGL2
— splats — fps
Splat size
Drag to orbit · scroll to zoom · right-drag to pan
Decoding splat model…

Every splat is an oriented 3D Gaussian — position, color, opacity and a covariance — solved so the set reproduces the input photographs. Sorted back-to-front and alpha-blended live; positions and covariance are quantized to keep the page light.

From a handful of photos to an interactive reconstruction

This public building example demonstrates the reconstruction workflow. Photographs go in; a photoreal model you can orbit, zoom, and pan around comes out.

Step 1

Capture

64 overlapping photos around the building, using perspective imagery rather than a depth scanner.

Step 2

Solve cameras

Feature matching + structure-from-motion estimate the 3D pose of every shot.

Step 3

Optimize splats

A dense field of 3D Gaussians is fit to the images until they re-render the scene photorealistically.

Step 4

Review

A view-consistent model for interactive orbit, zoom, and pan.

3D Gaussians (this model)
Source photographs
3072×2304
Pixels per photo
100%
Camera poses recovered

The input — ordinary photographs

A sample of the source frames. Nothing here is a render: each is a real photograph the pipeline matched against its neighbors to place in 3D space.