Gaussian splatting vs photogrammetry vs NeRF, in plain English

If you’ve been pitched 3D capture by more than one studio, you’ve probably had three different sales decks each describing a different technique as the future. This is an attempt to explain what they actually do, why they look the way they do, and which one you should care about for property work.

The short version: photogrammetry is the old reliable. NeRF was a research darling that never made it to mainstream property use. Gaussian splatting is what most professional studios are now shooting with. Here’s why.

Photogrammetry

The original. Take dozens or hundreds of overlapping photographs of a space, run them through software that figures out where the cameras were and what the surfaces look like, and out comes a 3D mesh model.

Matterport, broadly, sits in this family. So do most of the older 3D scanning workflows. The output is a polygon mesh draped in photo textures.

What it’s good at: producing a measurable 3D model. You can extract floor plans, dimensions and volume from a photogrammetry capture with confidence. RICS-grade floor plans for valuations come from this kind of capture.

What it’s bad at: photoreal rendering. The mesh approximates surfaces, the textures are painted on, and the result looks like a video game from 2010 rather than the actual room. It also struggles with reflective surfaces (mirrors, glossy floors, glass), thin objects (chair legs, curtains, plants) and anything that moves between captures.

If your priority is dimensioned floor plans, photogrammetry is still the right answer.

NeRF (Neural Radiance Fields)

The thing that got everyone excited in 2022. NeRF is a deep learning technique that learns a 5D function representing a scene’s volume and colour from a set of input photos. The output is a neural network you can render from any viewpoint.

The results were striking when they first emerged. Photoreal renders, smooth motion, handling of reflections and translucency that photogrammetry couldn’t touch.

The problem was rendering speed. A NeRF render of a single frame could take minutes on a high-end GPU. Real-time interactive viewing in a browser was effectively impossible.

NeRF basically didn’t make it into commercial property work. By the time the rendering engineers had it nearly fast enough, splatting arrived and obsoleted it.

Gaussian splatting

The current state of the art. Published as a research paper in 2023, in commercial use within months, mainstream by mid-2024.

The technique represents a scene as a cloud of millions of “Gaussians” — soft, blurry, oriented blobs that each carry colour, opacity and shape. Renders by projecting all of those Gaussians onto a 2D image and blending them together. The trick is that doing this is fast enough to run in a browser at 60fps, and the visual quality is closer to a photograph than anything that came before.

What it’s good at: photoreal interactive walkthroughs that run on phones. Reflections look right. Plants and chairs and curtains look right. Light spills look right.

What it’s bad at: precise dimensions. A splat doesn’t have a defined surface in the way a mesh does, so extracting an accurate floor plan requires a separate step (we re-render a vector floor plan from the splat geometry as a £75 add-on; some studios skip this). It’s also computationally heavier to capture and process than photogrammetry, which means studios charge a bit more for it.

For property marketing, splatting is now the right answer for any listing where the visual experience matters more than the dimensioned floor plan. That’s most listings above £500k, virtually all premium stock, and almost all developments.

Which one wins for property work in 2026

If you only care about a clean RICS-grade floor plan and a basic 3D walkaround for buyer self-qualification, photogrammetry through Matterport is fine.

If you want the tour itself to do marketing work — to make a buyer slow down, want to come and view, talk about the property to their partner that evening — splatting is the right answer.

NeRF is a research dead end for commercial property work. If a studio pitches you on NeRF in 2026, they’re either using it as a marketing label for splatting (which is legitimate, the underlying maths overlaps) or they’re behind the curve.

What changed at Zillow in 2025 and why it matters

In mid-2025, Zillow added native Gaussian splat support to their listings platform in the US. CoStar, Realtor.com and DJI followed within months. By the start of 2026, splat tours were no longer a novel feature, they were a category buyers expected to see on premium listings.

The UK market is six to twelve months behind the US on most property tech adoption curves. Splats are now common in the US, rare in the UK. That’s the window agencies should be thinking about. Two years from now, having a splat tour on a listing will be table stakes. Today, it’s still a meaningful pitch differentiator. We’ve written about where the cost lands and what the marketing benefits actually look like.

The technical answer for the engineers in the room

For anyone wanting more depth: a Gaussian splat is essentially a sparse representation of a radiance field. It throws out the neural network NeRF used and replaces it with a differentiable rasteriser that projects 3D Gaussians directly to screen. The maths is closer to traditional rasterisation than to deep learning, which is why it runs so fast.

Capture is similar to photogrammetry — a set of overlapping photos taken from many angles. The processing pipeline is what’s different. Splat training takes maybe four to twelve hours on a single high-end GPU for a typical residential interior. Output is between 200MB and 1.5GB of splat data, which the viewer streams progressively over a normal internet connection.

If you want to see what the format looks like for one of your listings, send us a brief at /contact. The first capture is on us for any agency that hasn’t worked with us before.