VOLUMESign up
Under the hood

How a splat is made.

The technique behind every VOLUME tour, in about four minutes. Nothing here is proprietary — the paper is public, the math is public, the reference code is on GitHub. We build the pipeline around it so you don't have to think about any of this.

Roan Priatel· Founder, VOLUMEUpdated 4 min read
Pipeline · four stagesphone → tour
  1. 01
    Video
    ~3 min · phone
  2. 02
    Point cloud
    structure-from-motion
  3. 03
    Gaussians
    ~2M blobs · trained on GPU
  4. 04
    Renderer
    Spark · WebGL 2 · 60 fps
Point cloud
Gaussians
Render
01

Frames become a point cloud.

Your walkthrough is a few thousand still frames of the same room from many angles. Structure-from-motion algorithms compare features between frames — the corner of a doorframe seen from two positions — and triangulate a sparse cloud of 3D points that matches the geometry of what you filmed.

This part isn't new. Photogrammetry has done it for decades. It's the input to what comes next.

02

Points become anisotropic Gaussians.

The Kerbl 2023 paper replaces the sparse points with elliptical Gaussians — soft, oriented blobs of colour that carry a position, a rotation, a scale, and a set of view-dependent colours. The scene is now a cloud of about a million to five million of these blobs. No mesh, no explicit surface, no texture atlas.

The reason this matters: the blobs are differentiable. A GPU can render a scene, compare it pixel-for-pixel to your input frames, and nudge the blobs' positions and colours to make the render match the reality. Iterate 30,000 times and you have a scene that photographs identically to what you shot.

03

Blobs render in a browser.

The output file is about 14 MB compressed. A WebGL 2 rasteriser — Spark, from Niantic — sorts the blobs front-to-back per frame and paints them. On a mid-range phone this runs at 60fps with no plugin, no headset, no download. On desktop it's indistinguishable from a rendered game engine.

04

Why this changes property media.

A splat is the first form of 3D interior media that (a) can be built from ordinary phone video, (b) fits in the size budget of a modern web page, and (c) renders in every browser. The previous options either required specialised camera hardware ($3,400 for a Matterport Pro3), required proprietary players (Matterport iframe, Zillow 3D Home), or both.

VOLUME's bet is that removing all three constraints collapses the market: what used to be a top-of-portfolio feature becomes a per-listing default. Which is what walkable listings are supposed to be.

Kerbl, B., Kopanas, G., Leimkühler, T., & Drettakis, G. (2023). 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics (SIGGRAPH 2023), 42(4), Article 139. Reference implementation at inria-lognormals/gaussian-splatting on GitHub. Spark, the WebGL 2 rasteriser we use, is MIT-licensed and maintained by Niantic.

You never have to touch any of this. That's the point.