Matterport vs Gaussian Splatting vs Photogrammetry in 2026: how we choose

For most briefs in 2026 the answer is unromantic: use Matterport when the deliverable is a navigable space someone has to walk through in a browser, use Gaussian splatting when a single object or interior has to look photoreal at film fidelity, and use photogrammetry when you need watertight, measurable geometry you own as files. They are not competitors. They are three tools that happen to start from the same act — pointing a camera at something real — and diverge completely after that.

We capture heritage sites, dealer showrooms and product for a living, and we have shipped all three pipelines into production. This is the decision rule we actually use, the tradeoffs that drive it, and where each one has burned us.

Matterport vs Gaussian splatting vs photogrammetry: what each one actually is

The terms get used interchangeably by clients and almost never mean the same thing. Definitions first, because the whole decision turns on them.

Photogrammetry reconstructs explicit 3D geometry — a triangle mesh with a texture map — from overlapping photographs. The output is a model: an .obj, .glb or .fbx with a UV-mapped texture. It is measurable, editable, and renders in any engine. It has been the archival standard for two decades.

Gaussian splatting (3D Gaussian Splatting, 3DGS, the 2023 technique from INRIA) does not build a mesh. It represents a scene as millions of overlapping translucent 3D blobs — Gaussians — each carrying position, scale, rotation, opacity and view-dependent colour. Rendered in real time, it reproduces reflections, soft edges and volumetric haze that a mesh physically cannot. The output is a .ply or .splat point cloud, not geometry you can put a tape measure on.

Matterport is not a representation, it is a product. A Pro2 or Pro3 camera captures depth and panoramas; Matterport's cloud stitches them into a hosted "dollhouse" walkthrough with measurement tools, floor plans and an embeddable viewer. You are buying a deterministic pipeline and hosting, not a file format.

So the real comparison is one open mesh format, one open radiance-field format, and one closed hosted service. That framing already tells you most of what you need.

The tradeoffs that actually decide it

Here is the comparison we keep in our heads on a capture call. The numbers are from our own jobs on OVH-hosted delivery behind Cloudflare, not vendor brochures.

• Visual fidelity. Splats win, decisively, for reflections, foliage, glass and atmosphere. Photogrammetry wins for crisp, well-lit matte surfaces — stone, plaster, machined metal. Matterport is the lowest fidelity of the three and does not pretend otherwise; it optimises for legibility, not beauty.
• Capture time on site. Matterport is fastest for whole buildings — a 400 m² floor in roughly 90 minutes of sweeps. Splats need disciplined photo or video coverage, typically 200–600 frames per scene. Photogrammetry is the slowest and least forgiving: it punishes any gap in coverage with holes in the mesh.
• File size and web delivery. This is where teams get surprised. A raw splat scene is routinely 150–600 MB before optimisation; we compress to roughly 30–80 MB with SOG/SPZ-style packing and still budget carefully on mobile. A web-ready photogrammetry .glb lands at 5–40 MB after decimation and Draco compression. Matterport ships nothing to your server — it streams from their cloud, so your page weight is just an iframe.
• Ownership and longevity. Photogrammetry and splat files are yours forever; you host them, you archive them, they outlive any vendor. Matterport tours live on Matterport's infrastructure and stop existing the day the subscription lapses.
• Measurability. Photogrammetry and Matterport both give real-world measurements. Splats do not — there is no surface to measure, only a field of blobs.
• Cost shape. Matterport is a recurring subscription per active space. Photogrammetry and splats are a one-time capture-and-process cost plus whatever hosting you already pay (for us, effectively zero marginal — it sits on the same OVH box behind Cloudflare).

Our decision rule

We do not pick by what is newest. We pick by what the brief has to do. The rule is three questions, in order.

1. Does the user need to navigate a whole space? If yes, and it has to be reliable, indexable and walk-throughable on a five-year-old phone, it is Matterport. A walkthrough is a UX problem before it is a rendering problem, and Matterport's viewer has solved that UX for years.
2. Is it one hero subject that has to look stunning? A single object, a single room, a façade — where reflections and light are the point — it is Gaussian splatting. This is the only technique that makes glass and polished surfaces read as real on a screen.
3. Does someone need to measure it, edit it, or keep it for decades? Conservation records, as-built documentation, anything that must survive a vendor or feed a CAD or game pipeline — photogrammetry, because you own watertight geometry as a file.

If two answers are yes, you have two deliverables, not one tool to compromise with. That is the single most expensive mistake in this field: trying to force one capture to serve a walkthrough and an archive and a hero render. We quote them separately because they are separately captured.

How this played out: Convento do Beato and Castelo de Castro Marim

Two heritage jobs make the rule concrete.

At the Convento do Beato in Lisbon, the brief was a navigable record of the event spaces — clients wanting to walk the halls before booking, plus a legible plan of how rooms connect. That is question one, unambiguously. We scanned it with Matterport: fast on site, a dollhouse view that sells the volume of the space, an embed that loads on anything. Nobody needed sub-millimetre geometry; they needed to understand the place in thirty seconds. Matterport was the boring, correct answer.

At the Castelo de Castro Marim, the requirement was different. The value was in the weathered stone, the irregular masonry, the light raking across a wall at the end of the day — and in a record durable enough to outlive any platform. We did not want a tour; we wanted fidelity and ownership. So the hero captures went to Gaussian splatting for the look, with photogrammetry on the structural detail where measurable, archive-grade geometry mattered. Files we host ourselves, on our own stack, citable and permanent.

Same studio, same week of planning, opposite tools — because the briefs were opposite. The castle could never have been a Matterport tour, and the convento walkthrough would have been absurd to deliver as a 400 MB splat nobody could navigate.

Where the three are heading in 2026

The honest state of the field:

• Splats are maturing into the web, not just demos. Compression formats and WebGL/WebGPU viewers have made them deliverable, and smartphone capture (Scaniverse and friends) is genuinely usable now. The remaining problem is delivery discipline — an unoptimised splat will wreck a mobile page.
• Photogrammetry is the quiet workhorse. No hype cycle, no change in what it is for. It remains the only output that is simultaneously measurable, editable and yours.
• Matterport's value is the part nobody tweets about: deterministic hosting and a viewer that just works. Their newer geometry is better, but we still reach for it for the same unromantic reason — it ships the walkthrough brief every time.

The product we are all waiting for is the one that fuses splat-quality rendering with deterministic, owned hosting. The day someone ships that cleanly, our rule gets simpler. Until then, three tools, three jobs.

The short version

Matterport for navigable spaces that must be reliable and indexable. Gaussian splatting for a hero subject that must look photoreal. Photogrammetry for measurable, editable geometry you own and archive. Pick by the job the deliverable has to do, not by what launched most recently — and when the brief asks for two of those at once, that is two deliverables, not one tool stretched thin.

If you have a space, an object or a heritage site that needs capturing and you are not sure which of the three it is, that conversation is exactly the part we are good at. Tell us about the brief.