Splats vs. the alternatives

A Gaussian splatA single anisotropic 3D Gaussian, an ellipsoid with position, covariance, opacity and view-dependent colour, used as the basic rendering primitive of a splat scene. Millions of these together, optimised against training images, make up the scene. isn’t the only way to turn photos of a real place into a 3D asset, and it isn’t automatically the right one. NeRFs, photogrammetry meshes and raw point clouds all solve the same underlying problem, recovering a scene from images or scans, with different trade-offs around speed, geometric accuracy and what the output is actually good for downstream. Picking between them is a project decision, not a technology preference.

Vs. NeRF: same field, different representation

Both a splat and a NeRFNeural Radiance Field: a scene representation that encodes colour and density as the weights of a neural network, rendered by marching rays through the volume and querying the network at each sample point. are ways of representing a radiance fieldA function that maps a 3D position and viewing direction to a colour and density or opacity, describing how a scene looks from any viewpoint. Both NeRFs and Gaussian splats are representations of a radiance field, one implicit, one explicit.: a function describing what colour and opacity you’d see from any point looking in any direction. The 2020 NeRF paper encoded that function implicitly, as the weights of a small MLP, recovered by marching rays through the volume and querying the network at each sample. That representation trains to a very compact file and generalises well to under-sampled regions, but rendering it means a network query per sample per ray, which is inherently slow. 3D Gaussian Splatting, published three years later, keeps the same target (a radiance field) but makes the representation explicit: a list of Gaussians a GPU can project and rasterise directly. You trade a larger file for the ability to render in real time, which is why splats displaced NeRFs for interactive delivery almost as soon as they appeared, even though NeRF variants can still win on raw reconstruction quality in some sparse-view or reflective-surface cases.

Vs. photogrammetry mesh: realism vs. measurable geometry

Photogrammetry (Structure-from-Motion plus multi-view stereo, the same COLMAP-style pipeline that seeds a splat’s training) builds an explicit triangle mesh with a texture, and it does so through triangulation: a mathematically grounded process tied to the actual physical geometry of the scene. That’s the mesh’s real advantage. The geometry carries metric accuracy you can measure, export to CAD, or bring into a BIM or GIS workflow, which is exactly why surveying, manufacturing and much of architecture still default to it. A splat scene sacrifices that measurability for visual fidelity: fine detail, reflections and soft or thin geometry (foliage, railings, glass) render convincingly in a splat in a way meshes tend to smooth over or artefact on, sometimes described as a “melted” look on complex scenes. If a deliverable needs to be measured, mesh it. If a deliverable needs to be walked through and look real, splat it.

Vs. point clouds: structure vs. raw samples

A point cloud, whether from LiDAR or from photogrammetry’s sparse or dense stereo output, is closer to raw data: a position and a colour per sample, with no shape, no opacity and no view-dependence. It’s often the input a splat’s training starts from (COLMAP’s sparse reconstruction seeds the initial Gaussians) rather than a competing deliverable in its own right. Point clouds are excellent for exactly the tasks they were designed for: measurement, registration and machine processing, where every sample means something numerically. They render as flat, gappy dots rather than a continuous surface, though, which is why they’re rarely used as the final visual product outside of specialist tools.

Where each one actually wins

None of these has made the others obsolete, and that’s unlikely to change soon: in 2025, Khronos and the OGC began work on bringing Gaussian splats into the glTF standard precisely because the industry expects splats to sit alongside meshes and point clouds as an interoperable asset type, not replace them. Pick a NeRF when you need the smallest possible file and can tolerate offline rendering. Pick a mesh when the deliverable needs to be measured, exported or built on top of. Pick a point cloud when you need the raw registered samples for surveying or as training input. Pick a splat when the deliverable is meant to be looked at, in real time, in a browser, and looking real matters more than being measurable.

In practice

On a site capture, the honest answer is often “more than one of these.” The same COLMAP run produces the sparse point cloud that seeds Gaussian training and, with an MVS pass, feeds a reference mesh for massing checks in an ARB-style review. The splat becomes the client-facing walkthrough in PlayCanvas; the mesh stays in the modelling tool for anything that needs a dimension pulled off it. Treating these as one pipeline with several outputs, rather than choosing a single winner up front, is what keeps a capture useful for more than the demo it was shot for.