Gaussian splats vs photogrammetry: two outputs from one capture

A Gaussian splat is a fast, photoreal way to render a place. It captures the look of a property — the soft edges, the reflections, the way light sits on a surface — and it renders smoothly enough to walk through on a phone. That makes it the right output for the experience: the thing you drop into a proposal so a customer can explore their own roof, or hand to a stakeholder who wants to look around without learning any software.

What a splat is not is a measurement substrate. Its representation is a cloud of fuzzy, view-dependent blobs optimized for appearance, not a clean surface you can put a tape measure on.

Photogrammetry takes the same images and solves for explicit geometry — a mesh, a dense point cloud, and an orthomosaic. These are the outputs you measure on: roof areas and pitches, distances, volumes, the deliverables a trade actually bills against. They're less glossy than a splat and that's the point; they're built to be correct, not just convincing.

Because both come from the same set of images, you don't run the capture twice. One trip to the site produces the walk-through scene and the measurable model side by side.

Generating both from one capture only works if the reconstruction is yours end to end — the structure-from-motion that recovers camera poses, the splat trainer, and the dense multi-view stereo that builds the mesh. Stitching together a vendor API for one and a different vendor for the other reintroduces exactly the seams that make capture painful.

The honest caveat travels with the geometry: a measurement is only as good as the capture that produced it. A quick uncontrolled pass gives you a preview-grade number; a survey-grade figure requires the RTK or ground-control inputs that anchor the model to real-world coordinates. The platform's job is to render both outputs and label the accuracy honestly — never to imply a precision the capture didn't earn.