3D Gaussian Splatting (3DGS) is one of the most talked-about technologies in reality capture right now, but the conversation often misses what matters most to practitioners. This article draws on Splat School, a ten-lesson field guide for the geospatial community, to explain what 3DGS actually is, where it excels, where it struggles, and why the strongest workflows treat it as a complement to LiDAR rather than a replacement.
Not a Photo, Not a 3D Model: What Gaussian Splatting Actually Is
If you work in LiDAR and reality capture, you have probably seen a Gaussian Splat. A photorealistic fly-through of a building interior, a heritage ruin you can walk through in a browser, a hotel lobby that looks more convincing than a photograph. The question worth asking is: what is it, technically?
3DGS starts with a point cloud, built either through Structure from Motion or directly via LiDAR. Using that point cloud as a scaffold, the software places millions of tiny oriented ellipsoids, called Gaussians or splats, through the scene. Each splat stores its position, size, shape, opacity, and view-dependent colour through a mathematical technique called spherical harmonics. The result can reproduce how surfaces actually look from any angle, because it learned that behaviour directly from photographs.
The rendering quality is genuinely different in kind from photogrammetry meshes. The way light plays across a polished floor, the soft shift in tone as you move past a window, reflections and surface sheen, all emerge naturally from the spherical harmonic encoding rather than fixed textures or post-hoc lighting calculations. For visualisation in hospitality, heritage, events and stakeholder engagement, 3DGS is hard to match.
The Accuracy Problem: Why 3DGS Cannot Replace a Measured Survey
Gaussian Splatting is optimised for visual quality. Geometric precision is a by-product, not the goal.
A Gaussian splat is not a surface. It is a soft, semi-transparent ellipsoid that fades to nothing at its edges. There is no hard boundary where the wall ends and the air begins. The optimisation decides the size and shape of each splat based on what looks best visually. Ask a 3DGS model for a wall dimension and you are asking it to do something it was never designed to do.
In my own testing on room-scale interior captures, 3DGS reconstructions have shown mean geometric errors of 7 to 8 centimetres. Results will vary with scene complexity, capture density, and scale. LiDAR sub-centimetre precision, by contrast, is well documented in controlled survey conditions. Reflective and transparent surfaces compound the problem further for 3DGS, though it is worth noting that LiDAR is not immune either, producing either absent or erroneous returns rather than reliable geometry. Scale is another issue: without physical reference points, a model can be internally consistent but dimensionally wrong.
For a measured survey you would submit to a planning authority, 3DGS alone is not the right tool. That is not a criticism, it is an accurate description of what the technology is for.
The Georeferencing Challenge: Four Problems That Need Solving Together
Placing a 3DGS model on a real-world map is more involved than it appears. There are four distinct problems, and you need to solve all of them.
First, the model has no idea where it is: no GPS reference, no north, no coordinate system of any kind.
Second, scale is not automatic: without survey control points at capture, the model cannot know its internal unit equals one metre.
Third, and this is the one that catches practitioners from photogrammetry or point cloud workflows: every part of the splat must move together. A splat is not a point. It stores its rotation through a covariance matrix and its view-dependent colour through spherical harmonic coefficients. Transform only the XYZ positions and the ellipsoids point the wrong way, colour fires in the wrong direction, and fidelity collapses.
Fourth, large coordinate values destroy precision. 3DGS stores coordinates in 32-bit floating point, around 7 significant digits. A British National Grid easting such as 333,456.782 uses six digits before the decimal point, leaving almost nothing for fine spatial detail. Splats centimetres apart round to the same value, producing a model that shimmers and jitters. This is catastrophic cancellation, and 3DGS is fully exposed to it.
The fix: do all transformation work in a local coordinate frame with small numbers, offset from a known origin, and convert to world coordinates only at the very last step.
Transforming a Splat Correctly: The Full Parameter Set
When merging, aligning, or geolocating a 3DGS model, the transformation must be applied to three things simultaneously: the 3D position, the covariance matrix defining each ellipsoid’s shape and rotation, and the spherical harmonic coefficients encoding view-dependent colour. Miss any one and the visual result will be wrong regardless of whether the geometry is correct.
For practitioners working with open tools, Splat-Transform by PlayCanvas handles translate, rotate, and scale across the full Gaussian parameter set. It merges multiple splat files and reads and writes all major formats including .ply, .splat, .ksplat, .spz and .sog. With survey control points providing real-world coordinates, apply the transformation correctly, work in local space throughout, and the result is a photorealistic model placed on a map to the accuracy of your control points.
Gaussian Splatting Versus Photogrammetry: Complementary, Not Competing
The question that comes up regularly: we already have photogrammetry, so why do we need this?
Photogrammetry produces measurable 3D surfaces, meshes and point clouds. It is well-integrated with CAD and BIM workflows and produces deliverables that surveyors, engineers and planning authorities accept. It handles thin structures, hard-edged geometry and large outdoor environments more reliably than 3DGS currently does. When you need accurate dimensions, it remains the right tool.
Gaussian Splatting brings photorealism, speed of capture, and visual immersion that photogrammetry meshes rarely match. Modern pipelines can process a room-scale capture in under two hours on good hardware. Real-time viewing on consumer devices is increasingly achievable. For sales, marketing, heritage engagement and event planning, 3DGS visual quality is hard to beat.
The strongest workflows treat them as complementary datasets: LiDAR or photogrammetry for the measured survey and engineering deliverables, Gaussian Splatting for the visual layer that clients connect with immediately. An increasing number of capture devices now output both survey-grade point clouds and 3DGS from a single site visit.
Where the Technology Is Heading: Standards, 4DGS, and the Digital Twin
The most significant developments are infrastructure-level. The Khronos Group is ratifying a native Gaussian Splatting extension to glTF, the most widely adopted 3D format on the web. OpenUSD version 26.03 brings splats into professional VFX and animation pipelines as a first-class asset type, with Nuke 17 already shipping native support. Both are infrastructure-level changes that will be felt for years.
4D Gaussian Splatting (4DGS) adds time to the equation. Framestore used it to deliver around 40 shots in Superman, capturing actors on a 192-camera stage and giving directors the ability to reframe and reshoot entirely in post-production. The implications for construction progress monitoring and heritage recording are beginning to be understood.
At the BIM and GIS intersection, georeferenced 3DGS models are becoming a practical interface layer for facilities management and digital twins with the photorealistic surface through which structured building data becomes navigable.
Key Takeaways for Geospatial Practitioners
- 3DGS produces photorealistic, real-time explorable models from photographs, built on a point cloud scaffold derived from LiDAR or Structure from Motion.
- It is optimised for visual quality, not geometric precision. In room-scale interior testing, mean geometric errors of 7–8 cm have been observed; results vary with scene complexity, capture density, and scale. It is not a replacement for a measured survey.
- Georeferencing requires transforming positions, covariance matrices, and spherical harmonic coefficients together. Transforming positions alone collapses visual fidelity.
- Work in a local coordinate frame during transformation. Large coordinate values exhaust 32-bit precision and introduce visible artefacts.
- The strongest practitioners treat 3DGS and LiDAR as complementary. One capture, multiple deliverables, each suited to its purpose.
- glTF and OpenUSD standardisation will define how 3DGS integrates into BIM, GIS and FM workflows for years to come.
Splat School: A Field Guide to Gaussian Splatting is available as a free PDF download at https://www.cr8ive.ai
About the Author
Mike Devers
Director, Cr8ive Media UK Limited
Mike Devers is a spatial data capture and AI visualisation specialist with a 40-year background in enterprise technology, including senior roles at Dell UK and EMEA. Through Cr8ive Media, he works with clients in hospitality, heritage, and events, combining LiDAR capture, Gaussian Splatting, and AI visualisation tools to produce photorealistic spatial experiences. He is the author of Splat School, a ten-lesson field guide to Gaussian Splatting, and runs the Workflow Accelerator training programme for 3DGS practitioners.
Website: cr8ive.uk
LinkedIn: linkedin.com/in/mikedevers
