3D Gaussian Splatting for Visualization: Where Photorealistic Reality Capture Actually Helps
Gaussian splatting turns scan and photo data into photorealistic, navigable scenes. Here is where it earns its place in an AEC or heritage workflow, and where LiDAR geometry still needs to lead.
Gaussian splatting has become the most talked-about visualisation technique in reality capture since photogrammetry went mainstream. It is genuinely impressive — and genuinely easy to misapply. This article separates what it is good at from what it cannot do, and describes how we combine it with measured geometry rather than treating it as a replacement.
If you have seen a scan walkthrough recently that looked less like a model and more like a photograph you could walk through, it was probably Gaussian splatting. The technique emerged from academic computer graphics in 2023 and has since spread through architecture, heritage and infrastructure workflows fast enough that clients now ask for it by name.
The enthusiasm is justified. The confusion is not.
What Gaussian splatting actually is
Traditional reality capture produces one of two things. Photogrammetry produces a mesh — a surface of triangles with a photographic texture pasted on. Terrestrial LiDAR produces a point cloud — millions of accurate measurements rendered as coloured dots.
Gaussian splatting produces neither. It represents the scene as a very large number of soft, semi-transparent ellipsoids — "splats" — each with a position, a size, an orientation, a colour, and an opacity that changes with viewing angle. Because the representation is view-dependent, highlights, reflections and fine surface detail behave much more like the real thing than a fixed texture can.
The practical consequences that matter on a project:
- Photorealism that survives close inspection. Foliage, rust, worn stone and reflective metal look correct rather than approximated.
- Interactive frame rates. Unlike dense meshes, splats render efficiently enough for smooth navigation on ordinary hardware.
- Forgiving capture requirements. Splats can be trained from overlapping photographs alone, so a phone or drone capture can produce a compelling scene.
The mistake to avoid
The mistake is treating a splat as a measurement product.
Splats are optimised for appearance, not for dimensional truth. There is no reliable way for a project manager to pull a trustworthy dimension, section or clash check out of a splat scene. If you need to know whether a beam is 12 mm out of position, that answer lives in the registered LiDAR point cloud, and it needs to stay there.
We treat the two as complementary layers over the same site:
| Layer | Source | Used for |
|---|---|---|
| Measured geometry | Terrestrial LiDAR (FARO Focus S 350, ±1 mm ranging) | Dimensions, sections, deformation checks, Scan-to-BIM |
| Visual fidelity | Gaussian splats trained on fused photos and scan data | Stakeholder review, public engagement, training, marketing |
Both layers are georeferenced to the same control network, so a reviewer can switch between "looks real" and "is precise" without leaving the same coordinate space.
The pipeline, end to end
A production splat deliverable is less glamorous than the demos suggest.
- Capture. Overlapping imagery from a UAV, handheld or tripod-mounted camera, with sufficient baselines and coverage. Where the site is complex or occluded, we capture LiDAR in parallel for geometry and registration control.
- Structure-from-Motion. Photos are aligned into camera poses. Ground control points tie the reconstruction to survey coordinates so the result is not just pretty but placed correctly.
- Training. The optimisation step that fits splats to the imagery. This is GPU-bound and scales with scene size and target quality.
- Cleanup and optimisation. Pruning stray floaters, reducing splat counts where detail is not needed, and tuning for the delivery platform.
- Delivery. Export to the format the viewer needs, then publish — web stream, offline viewer, or both.
Steps three and four are where most of the effort goes, and where unrealistic expectations usually originate.
Where it pays off commercially
Splat visualisation tends to justify itself in four situations:
- Stakeholder and community review. Heritage interpretation, public consultation and planning committees respond very differently to a photoreal walkthrough than to a grey mesh.
- Tender and bid material. A navigable scene of an existing asset communicates site conditions in a way drawings cannot.
- Training and remote support. Operators and maintenance crews can rehearse access to plant or heritage interiors safely, before arriving.
- Marketing and documentation. Museums, tourism bodies and asset owners use splat scenes as a durable visual record of a condition at a point in time.
Honest limits
Three constraints deserve to be stated plainly before anyone scopes a splat deliverable.
File size and hardware. Raw splat scenes can run from hundreds of megabytes to several gigabytes. Delivering them implies a viewer strategy: level-of-detail streaming for web, or a specified workstation for offline review.
Not CAD-native. Splats do not open in Revit, AutoCAD or a GIS as usable geometry. If downstream teams need to model, the point cloud and BIM model are the products; the splat sits alongside.
Capture discipline still decides quality. Blurry imagery, thin overlap, moving objects or poor lighting produce artefacts that no amount of training fixes. The technique is forgiving, not magic.
How we deliver it
For most projects, we recommend the layered approach described above, with splats serving as the visual interface over measured LiDAR data. That gives clients the persuasive experience of a photoreal scene without sacrificing the dimensional rigour their engineers need.
If you want to explore the capture side in more depth, see our 3D laser scanning and reality capture service or read the companion piece on point cloud to BIM standards. For heritage applications specifically, HBIM documentation is usually the right starting framework.
Frequently Asked Questions
What is 3D Gaussian Splatting in simple terms?
Instead of representing a scene as triangles (a mesh) or individual measurements (a point cloud), Gaussian splatting represents it as millions of soft, volumetric blobs. Each blob carries position, size, orientation, transparency and colour depending on viewing angle. Rendered together, they produce a photorealistic, freely navigable scene at interactive frame rates.
Can Gaussian splats replace LiDAR for measurement?
No. Splats are a visual representation, not a measurement-grade geometry product. For dimensional accuracy, deformation checks or BIM production, the LiDAR point cloud remains the reference. The practical answer is to pair them: geometry from LiDAR, visual fidelity from splats.
What file formats do Gaussian splats come in?
Common deliverables include .ply (the de facto exchange format), plus optimised runtime formats such as .splat, .ksplat or SOG for web and mobile delivery. We scope the format to the viewing platform — desktop review, web streaming or offline kiosk.