Gaussian splats: the past, present and future of capturing reality
From photogrammetry to NeRFs to 3D Gaussian Splatting — how reality capture became real-time, and why a phone scan of a Jamaican cell tower now beats a survey crew.
For thirty years, turning the physical world into a digital model meant photogrammetry: hundreds of overlapping photos, hours of processing, and a mesh that looked like melted wax anywhere the algorithm lost confidence. It worked — for film studios and survey departments with time and budget. It never worked for the person who needed an answer today.
The first rupture came in 2020 with Neural Radiance Fields. NeRFs stopped trying to reconstruct geometry and instead learned a scene as a function: ask it what light arrives from any direction, and it answers. The results were startlingly photoreal — and unusably slow. Minutes per frame is a research demo, not an operations tool.
Then, in 2023, 3D Gaussian Splatting changed the economics. Instead of a neural network, a splat represents the scene as millions of tiny translucent ellipsoids — Gaussians — each with position, scale, colour and opacity, optimised directly against the source images. The decisive property is that they rasterise: standard GPU pipelines render them at real-time frame rates. Photoreal capture stopped being an offline process and became something you can walk through, live, in a browser.
That is the present, and it is why we build on splats. A vehicle-mounted or handheld scan — even a phone — becomes a navigable, photorealistic twin in hours. A tower estate in Jamaica can be captured by a local driver, splatted overnight, and inspected from Gothenburg the next morning: guy-wire anchors, corrosion, antenna azimuths, encroaching vegetation. The expert no longer flies to the asset; the asset flies to the expert. Research groups — Bell Labs among the leaders — keep pushing quality and compression, and every improvement drops straight into the pipeline.
The near future is splats plus semantics. A splat today is beautiful but naive — it knows colour, not meaning. The next layer assigns every Gaussian to an asset: this cluster is an antenna, that one a transformer, this discolouration is rust that wasn't there in April's baseline. Diffing two dated splats of the same site turns reality capture into change detection — the foundation of storm damage assessment, predictive maintenance, and bills of materials generated by agents instead of clipboards.
Further out, splats stop being photographs and become simulation substrates. Physics engines are learning to run against splat geometry directly: RF propagation for network planning, water flow, fire spread, crowd movement. When the photoreal model and the physics model are the same object, the digital twin stops being a dashboard and becomes a rehearsal space for reality.
The pattern behind all three eras is the same one we apply everywhere: work with what you already have. Photos you already take, drives you already make, cameras you already own — up to 99% of the value comes from connecting existing signals, not buying new sensors. Gaussian splatting is simply the first technology that makes that principle photorealistic.
One call is enough to know if we're a fit.