Skip to content

A nice explanation of Gaussian splatting

Published: at 08:36 AM

Gaussian splatting reconstructs 3D scenes from ordinary photos or video. You capture a subject from multiple angles, software tracks the camera positions, and a neural network trains millions of tiny 3D Gaussians — colored, transparent ellipsoids — to reproduce what the cameras saw. The result renders in real time: no mesh, no texture map, just millions of tiny brush strokes floating in space, each one carrying its own color, size, and transparency.

The technique emerged from a 2023 research paper and spread fast. Game engines, VFX pipelines, and phone apps adopted it within a year.

The workflow

Corridor Crew walks through the full pipeline — capturing footage, processing it, and rendering the final splats in real time. They’re a VFX studio, so the emphasis falls on where this fits alongside traditional CGI. The short version: it’s faster, cheaper, and the quality holds up.

Printing splats

Gaussian splats live on screens. They are volumetric data — light, color, density — with no solid surface to print. A project called DreamPrinting, now commercialized as Crysta, bridges that gap by translating radiance field data into instructions for full-color 3D printers, accounting for how light scatters through physical material. The translucency and density carry over into the print — not just a shell with a texture on it.