Gaussian splatting vs photogrammetry: what 10 practitioners say

At a glance

Two directions from the same photos

Olli Huttunen, who explains splatting as pointillism in 3D, describes both methods as starting from the same point cloud and then going in different directions: a mesh for physical uses such as 3D printing, or a splat model built to look good on a screen. A splat is millions of soft points trained until the rendered view matches the original photos – and, in Olli Huttunen’s words, it has no geometry.

That framing runs through the whole guide: the question is not which technology is better, but what the client will do with the result.

Where splats win: glass, water, wires and plants

Dylan Gorman of PilotByte names water surfaces and reflective glass as the biggest reason to choose a splat – both break in photogrammetry. Olli Huttunen adds that transparency is almost impossible with mesh structures. Arash Keissami of NIRA has seen splats handle thin wires and interiors remarkably well, and Barry Bassnett likes that vegetation on a splat shows the actual plants rather than a symbol.

The limit, per Olli Huttunen: splats look photoreal from a distance, but go soft when you get close to leaves or surfaces – and anything that moved during capture can leave ghosts.

Where photogrammetry wins: measuring

Every guest who measures for a living keeps measurement off the splat. Dylan Gorman switches to photogrammetry when a client needs to see cracks or take measurements. Arash Keissami warns that semi-transparent splats can be much larger than they look, so a click can land in the wrong place – NIRA adds mesh and matcap views for that reason. Michael Schoelkopf of NavVis describes the same split: show the splat, measure on the point cloud.

There is one important exception. Nikhil Sawlani of MultiSet AI uses splats for visual positioning – but only splats at metric scale, such as those generated with a LiDAR sensor underneath, or scaled manually. A splat from photos alone won’t do.

Capture: more angles, fewer (and smaller) images

Barry Bassnett is blunt that image quality decides everything – an out-of-focus capture makes a bad splat just as it makes a bad mesh – and recommends HDR bracketing so harsh sun doesn’t ruin the images. Olli Huttunen prefers overcast weather for the same reason.

The difference is in how the images are used. Olli Huttunen says a splat breaks apart at angles that were never shot, so capture broadly at several heights – but then train on a hand-picked subset, because images above 4K mostly slow the training down. Photogrammetry, by contrast, benefits from thousands of high-resolution images. Dylan Gorman’s water-tower test shows the trade-off: a quick splat was impressive, but the full photogrammetry capture still won on fidelity.

File size, the browser and VR

Both methods get heavy. Olli Huttunen says a model with 12 million splats runs to gigabytes, and that the compressed SOG format makes a splat about 95% smaller for the web. Arash Keissami says large photogrammetry models and splats alike crash phones, which is why NIRA renders on a server GPU and streams the result.

For speed of production, Alex Harvey of RiVR compares a photogrammetry room – thousands of images, processing, clean-up and retopology – with a Meta Hyperscape splat that gives the same VR experience in hours. Olli Huttunen and Alex Harvey both call a headset the best way to experience splats; Barry Bassnett is sceptical that VR is used much for digital twins.

Which deliverable for which client

Dylan Gorman asks clients one question up front: do you want to look at it, or measure it? A splat covers “it exists, I need to see it” documentation, site walkthroughs and real estate; photogrammetry covers inspection and measurement. Barry Bassnett suggests surveyors deliver a splat of the site alongside the 2D topographic survey, for architects, planners and the public.

Zlata Karpushina of FJ Dynamics sees splats as the layer that helps non-technical stakeholders understand a space – but only one step in a workflow, because visualisation alone solves half the problem. Arek Szadkowski of Esri adds the market view: early adopters still deliver splats as an add-on extra on top of what the client ordered, and tenders that specify them were almost non-existent as of October 2026.

Combine them – and will splats replace photogrammetry?

The practical answer from most guests is to combine. Olli Huttunen reprocesses old photogrammetry datasets as splats and mixes splat environments with mesh models. Barry Bassnett describes a splat as the visual “atlas” with a mesh or point cloud underneath for tagging and coordinates, and Bilawal Sidhu points to Apple’s hybrid of photogrammetry and splats for maps.

On replacement, the guests split. Barry Bassnett says photogrammetry is dying except for specific needs; Dylan Gorman says splats are not a replacement at all; Zlata Karpushina and Michael Schoelkopf see splats as a layer on top of measurable data. Arash Keissami’s view is the safest bet: it is still very early. In Esri’s ArcGIS Reality, per Arek Szadkowski, splats are simply one more export from the same alignment as orthos, meshes and point clouds – and the new Ground Splats from 360 mobile mapping take their geometry from LiDAR.

Questions

Can you take measurements on a Gaussian splat?

The guests advise against it for now: Arash Keissami (NIRA) says splats can be deceiving when measuring, and Michael Schoelkopf (NavVis) and Dylan Gorman measure on the point cloud or photogrammetry instead. Metric-scale splats, e.g. from a LiDAR-based scanner, are the exception Nikhil Sawlani (MultiSet AI) uses for positioning.

Why do Gaussian splats look better than photogrammetry on glass and water?

Splats are view-dependent and semi-transparent, so they reproduce reflections and transparency that meshes cannot – Dylan Gorman says water and reflected glass break every time in photogrammetry.

Can I make a Gaussian splat from my photogrammetry photos?

Yes – Olli Huttunen says the same image datasets can be reused, and older photogrammetry captures can be reprocessed as splats; train on a curated subset rather than thousands of images.

Will Gaussian splatting replace photogrammetry?

The guests disagree: Barry Bassnett expects splats to replace it for many applications, Dylan Gorman says it is not a replacement at all, and Zlata Karpushina (FJ Dynamics) sees splats as a visual layer on top of measurable data.