EN / SKBook a demo
The technology

Gaussian splatting, explained.

A plain-language guide to the technology behind Splatoo — what it is, how it differs from older 3D methods, and why it makes real places explorable on any screen.

What it is

Gaussian splatting captures a real place in 3D as millions of tiny, soft points of light — not flat polygons.

Each point is a "splat": a position in space, a colour, a softness that decides how it fades at the edges, a level of transparency, and a shape — it can stretch and tilt to hug the surface it sits on. On its own a single splat is just a fuzzy blob. Stack a few million of them and the original scene reappears, with true depth, fine detail, and the exact light and atmosphere of the place it was captured in.

Under the hood it's a radiance field — a model of how light leaves every point of a scene in every direction. The first radiance fields (NeRFs) hid that information inside a neural network and had to compute each pixel on demand. Gaussian splatting stores it out in the open, as the splats themselves. There's nothing to infer at view time, so it draws fast — fast enough to explore smoothly in a browser, on a phone, with no app to install.

A splat carries five things: position, colour (which can shift depending on the angle you look from), opacity, size and orientation. That last pair — being able to stretch and rotate — is exactly why splatting holds onto soft, reflective and fine detail that polygon meshes throw away.

How it works

The remarkable part is that you don't model anything by hand. You photograph or film the space, and an optimiser rebuilds it for you.

  1. Capture

    Walk the space with a camera or even a phone, covering it from many overlapping angles. A few hundred frames is often enough for a room; larger spaces simply need more coverage.

  2. Locate the cameras

    Structure-from-motion software studies the frames, works out where each photo was taken, and lays down a rough starting cloud of points to build from.

  3. Optimise the splats

    Every point becomes a 3D Gaussian. The system renders the scene, compares it against the real photos, and nudges each splat's position, colour, size and opacity to close the gap — millions of tiny corrections, repeated until the render is hard to tell from the photographs. Along the way it adds splats where the scene is still blurry and deletes ones that aren't earning their place.

  4. Render in real time

    To draw a frame, the splats are projected onto your screen, sorted by depth and blended front-to-back. No neural network runs at view time — which is the whole reason it hits smooth, real-time framerates on ordinary hardware.

Training takes minutes rather than the hours older neural methods needed. And because the result is explicit — a literal set of splats rather than a black-box network — the scene can be moved, recoloured, trimmed or combined with others. That editability is what we build on.

How it's different

Every earlier way of capturing reality trades something away. Gaussian splatting is the first that keeps photoreal quality and real-time speed at the same time.

Photogrammetry
Builds a textured polygon mesh. Accurate for hard surfaces, but heavy — and soft, reflective or fine detail like foliage and hair tends to break down.
NeRF
Stunning quality from a neural network, but historically slow to train and too heavy to render in real time, especially in a browser.
Gaussian splatting
Photoreal like NeRF, trains in minutes, renders in real time, and stays light enough to run on the open web.

It also clears the bar set by two older shortcuts. Raw LiDAR and point-cloud scans capture accurate geometry but look flat and lifeless without real lighting baked in. Plain video looks real but traps the viewer on a single fixed path. Splatting gives you both at once — the look of video with the freedom to move anywhere inside it.

The short version: it's the first method that looks like the real place and runs anywhere.

Strengths & limits

It's worth being straight about where the technique shines today and where it's still maturing.

Strong at

  • Rich static spaces — interiors, architecture, landscapes, single objects
  • Soft and fine detail — foliage, fabric, hair, smoke
  • View-dependent light — gloss, sheen and subtle reflections
  • Fast capture and instant, app-free viewing on any device

Still maturing

  • Moving subjects — people and traffic are best captured when the space is empty
  • Very large scenes — these need streaming and compression to stay light
  • Precise editing and relighting — improving fast, but not yet one click
  • Mirrors and glass — true reflections can confuse any capture method

Most of these are active research fronts moving quickly — and several are exactly where our capture and Unreal pipeline fill in the missing pieces.

Why it matters

A render shows someone what a space could look like. A scan lets them stand inside what it actually is. That difference is the whole point: people decide faster, trust what they're seeing more, and genuinely remember a place they've moved through rather than one they've only glanced at.

For a business, that turns into fewer wasted site visits, shorter sales cycles, and a single asset that works on a website, a sales call and an on-site screen alike. One capture, many places it can live.

And it's no longer experimental. Real-estate platforms have begun handing buyers splat tours, the technique has found its way into major film and visual-effects production, and capture apps are putting it in everyday hands. The industry calls this a "JPEG moment for spatial computing" — the point where 3D capture stops being a novelty and becomes ordinary.

Where it's heading

The field is moving unusually fast. Four fronts matter most for where this goes next:

  • Compression — scenes shrinking from gigabytes toward megabytes, so a full space loads like a web page.
  • Dynamic "4D" capture — recording a space as it changes over time, not just a frozen moment.
  • Relighting — dropping a captured space into new lighting, or matching it to a new environment.
  • On-device capture — turning a phone into the entire studio, from scan to shareable scene.

The direction is clear: spatial capture is becoming as casual as taking a photo — and as easy to share. We're building Splatoo for that world.

What we do with it

A raw scan is beautiful but passive. Splatoo turns it into something people can use.

We capture the place, clean it up, then layer a wayfinding system on top — points of interest, routes, services and content — so the space isn't just viewable, it's navigable. With our Unreal pipeline we can also bring in objects and scenes that don't physically exist yet, so reality becomes the starting point rather than the limit. The finished experience publishes straight to the web and to on-site kiosks, and quietly reports back on how visitors actually move through it.

One capture becomes a presentation, a sales tool, a guide and a touchpoint — all at once.

See it live

Live scene · 1.2M points · WebGL

Reading about it only goes so far. Open a real walkthrough in the demo archive and move through a captured space yourself — drag the points, switch scenes, and see how it feels in the browser.

References & further reading

The plain-language explanation above is grounded in the primary research and reporting below.

  1. Kerbl, Kopanas, Leimkühler & Drettakis — "3D Gaussian Splatting for Real-Time Radiance Field Rendering." ACM Transactions on Graphics (SIGGRAPH 2023) — the paper that introduced the technique. Project page · Reference code
  2. Inria — "Creating stunning real-time 3D scenes: the breakthrough of 3D Gaussian Splatting." A readable overview from the lab behind the method. inria.fr
  3. Mildenhall et al. — "NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis." ECCV 2020 — the neural radiance field work that splatting builds on and accelerates. Project page · arXiv
  4. Schönberger & Frahm — "Structure-from-Motion Revisited" (COLMAP). CVPR 2016 — the camera-pose and point-cloud step that seeds a capture. colmap.github.io
  5. Gaussian splatting — overview. A continually updated summary of the field, its variants and adoption. Wikipedia
  6. "Gaussian Splatting and the Infrastructure for Spatial Computing." Industry analysis behind the "JPEG moment for spatial computing" framing. fov.ventures
See it for yourself

Want your space captured like this?

We'll scan it, layer it, and hand you a walkthrough that runs anywhere.