From polygons to brushstrokes
Traditional 3D models are built from polygons — flat triangles stitched together to approximate surfaces. Walk through a photogrammetry mesh and you'll see jagged edges, missing details, and artifacts where the geometry couldn't keep up with reality.
Gaussian splatting takes a different approach entirely. Instead of polygons, it represents a scene as millions of tiny, soft 3D ellipsoids — think of them as translucent brushstrokes suspended in 3D space. Each one has a position, a size, a shape, a color, and an opacity. Together they reconstitute the world not as a hard surface, but as a volumetric cloud of light.
The result is something that feels less like a scan and more like stepping back into the place itself.
Four properties no polygon mesh can replicate
Gaussian splatting isn't just a different file format — it's a fundamentally different way of representing reality. Here's what sets it apart.
Fine Detail
Hair, texture grain, fabric threads — Gaussian splats preserve detail at a level polygon meshes simply cannot match. Millions of splats can cluster around high-frequency surfaces, producing photorealistic fidelity without any topology clean-up.
Fine Structures
Fences, rebar, scaffolding, vegetation, wire runs — geometry that frustrates photogrammetry because it has no closed surface. Gaussian splats just place ellipsoids wherever photons arrive, making thin structures appear in full fidelity.
Transparency
Glass, windows, water, mesh screens — surfaces that transmit light. Each Gaussian splat has its own opacity value, so semi-transparent materials render naturally. You can see through the glass and still see what's on the other side.
View-Dependent Reflections
Mirrors, polished floors, wet surfaces — materials whose appearance changes depending on where you are standing. Each splat stores spherical harmonic coefficients that encode how its color shifts with viewing angle, producing accurate reflections without any special material definition.
Transparency and reflections — together
The real test of any 3D capture system is a glass storefront at night — transparent surface, dramatic reflections from the street, neon signs bouncing off the floor. This scenario breaks polygon meshes completely.
Gaussian splats handle it naturally. The opacity channels let light pass through. The spherical harmonics encode the reflection as it actually appeared from each camera. The result is spatially accurate, visually faithful, and navigable in real-time.
- Glass facades rendered with accurate transmission
- View-dependent reflections shift as you move through the scene
- No special material passes or manual retouching required
How a Gaussian Splat is trained
A Gaussian splat isn't exported from a scanner. It's trained — like a neural network — through an iterative optimization loop that refines millions of tiny ellipsoids until the rendered output matches your photographs.
Capture
A camera circles the subject — a room, a site, an object. Dozens or hundreds of overlapping photos from different angles.
Structure from Motion
Software analyzes the photos to triangulate where each camera was positioned and build a sparse point cloud of the scene.
Splat Initialization
A Gaussian ellipsoid is seeded at each point in the sparse cloud — blurry, opaque, roughly the right color.
Gradient Descent
The AI renderer projects the splats onto virtual cameras and compares the result to the real photos. Parameters are adjusted thousands of times, nudging position, scale, rotation, color, and opacity toward photorealism.
Densification & Pruning
Splats that are too large get split. Splats in the wrong place are removed. The cloud tightens until the rendered view is nearly indistinguishable from the original photograph.
Gaussian splats materializing during training — from sparse noise to photorealistic 3D.
Gaussian Splats vs. Everything Else
An honest look at how Gaussian splatting compares to photogrammetry and LiDAR point clouds — including where the other technologies still win.
| Capability | Gaussian Splat | Photogrammetry | LiDAR Point Cloud |
|---|---|---|---|
| Visual Fidelity | |||
| Photorealistic output | Partial | — | |
| Fine detail (hair, fabric, texture) | — | — | |
| Thin structures (rebar, fencing, wire) | — | Partial | |
| Transparent surfaces (glass, water) | — | — | |
| View-dependent reflections | — | — | |
| Real-time browser rendering | Partial | — | |
| Shareable without desktop software | — | — | |
| Measurement & Engineering | |||
| Survey-grade dimensional accuracy | Partial | ||
| Direct depth measurement (no image matching) | — | — | |
| Volumetric & area calculations | — | ||
| Works in low-light / featureless environments | — | — | |
| Workflow & Interoperability | |||
| Editable mesh / CAD-BIM compatible output | — | Partial | |
| Scan-to-BIM workflow | — | Partial | |
| Industry-standard file formats (OBJ, FBX, LAS, E57) | Partial | ||
| Established software ecosystem | Partial | ||
| No specialist hardware required | — | ||
“Partial” indicates the capability is possible but with significant limitations or requires additional tooling.
Where Gaussian Splats are changing the game
Any industry that needs to communicate space, capture as-built conditions, or share immersive 3D — without specialist hardware.
Capturing the splat is just the beginning
Splat Labs turns raw Gaussian Splats into shareable, measurable, annotatable deliverables — in your browser, with no specialist software.
Host & Share
Publish splats to a secure, shareable link. Control who can view — public, private, or team-only.
Measure
Take precise point-to-point, area, and volume measurements directly inside the 3D model.
Annotate
Pin notes, documents, photos, and videos to any point in the 3D scene for team collaboration.
Build Tours
Create guided walkthroughs with auto-camera paths and a filmstrip navigator.
Overlay Data
Drape KML layers and orthomosaics over your splat for geospatial context.
AI Scene Redesign
Remove furniture, redesign spaces, or insert 3D objects with a text prompt.
Frequently Asked questions
01What is the difference between a Gaussian splat and a point cloud?
A point cloud is a collection of discrete XYZ coordinates — dimensionless dots in space. Each point has no size, no shape, no opacity. Gaussian splats are volumetric: each one is a soft 3D ellipsoid with orientation, scale on three axes, color, and transparency. The result renders as photorealistic imagery rather than a cloud of colored dots.
02How does Gaussian splatting compare to photogrammetry?
Photogrammetry reconstructs geometry as a closed polygon mesh and then drapes a texture photograph over it. It struggles with transparency, thin geometry, and view-dependent reflections because those require the renderer to understand volume and light direction — not just surface position. Gaussian splatting doesn't try to build a mesh. It represents the scene as a cloud of volumetric light sources, which naturally handles all of these edge cases.
03What hardware do I need to capture a Gaussian Splat?
You can capture a Gaussian Splat with any camera that produces overlapping photographs — including a smartphone. Purpose-built hardware like the PortalCam or XGRIDS Lixel L2 Pro produces higher density and better geometry, but the technology is hardware-agnostic. Splat Labs accepts splats from Polycam, Luma AI, DJI, and any other tool that outputs the standard .ply or .splat format.
04How long does training take?
A typical interior scene trains in 20–60 minutes on consumer GPU hardware. Purpose-built training servers or cloud training platforms can reduce this to minutes. Training time scales with scene complexity and the number of input photographs.
05Can I take measurements inside a Gaussian Splat?
Yes, with Splat Labs. Measurements are taken against the underlying point geometry that anchors the splats, giving sub-centimeter accuracy for most scenes. You can measure distances, areas, and volumes directly in the browser without any desktop software.
06Do Gaussian Splats work in VR or AR?
Real-time VR rendering is an active area of development. Splat Labs currently supports VR viewer mode in-browser. Native XR rendering (Quest, Vision Pro) is on the roadmap as GPU rasterization pipelines for splats mature.
07How large are Gaussian Splat files?
A typical interior scene with 1–3 million Gaussians is 200–600 MB in raw .ply format. Splat Labs applies compression and streaming so viewers load quickly in the browser regardless of file size.
08Can I embed a Gaussian Splat on my own website or in Zillow?
Yes. Splat Labs generates embeddable iframes that you can paste into any web page, Zillow listing, or content management system. Enterprise users can white-label the viewer with their own branding.