NeRF vs 3D Gaussian Splatting vs Photogrammetry vs LiDAR Scanning for Digital Twins: Accuracy, Speed, Cost and When to Use Each
A Gaussian splat of a pump room can look more real than any laser scan you have ever paid for, and still be the wrong thing to hand to a BIM modeller. That gap between looking right and measuring right is the whole story of nerf vs gaussian splatting for industrial and building digital twins. Most comparisons rank the four reality-capture methods on image quality. Owners of plants, fabs and hospitals care about something else: can the capture carry a tolerance, can it be versioned, and can a pipe be traced into a CAD or IFC model? This matters now because 2026 is the year splats became a standards citizen. OpenUSD v26.03 added a native Gaussian splat schema, and Khronos put a glTF splat extension into release candidate. You will leave with a mechanism-level comparison, worked numbers for accuracy and file size, a decision matrix by use case and a hybrid architecture that uses each method for what it is good at.
What this covers: how each representation works, metric accuracy and scale, processing and rendering speed, editability, CAD and BIM compatibility, file formats and standards status in November 2026, tooling, failure modes and a use-case decision matrix.
Context and Background
Reality capture for digital twins has had two incumbents for two decades. Terrestrial laser scanning (TLS) measures distance directly with time-of-flight or phase-shift lidar and produces a point cloud with survey-grade error budgets. Photogrammetry recovers geometry from overlapping photographs through Structure from Motion (SfM) and Multi-View Stereo (MVS) and produces a dense cloud and a textured mesh. If you want the physics of the first, read how lidar actually works, from photon budget to range noise.
Two newer methods came out of graphics research rather than surveying. Neural Radiance Fields (NeRF) were introduced by Mildenhall and colleagues in 2020 (arXiv 2003.08934). A NeRF is a multilayer perceptron (MLP) that maps a 3D position and a viewing direction to a density and a colour, and images are made by volume rendering along camera rays. 3D Gaussian Splatting (3DGS), from Kerbl, Kopanas, Leimkühler and Drettakis at SIGGRAPH 2023 (arXiv 2308.04079), swapped the network for millions of explicit, anisotropic 3D Gaussians that are rasterised rather than ray-marched. The paper’s headline claim is real-time novel-view synthesis, at or above 30 frames per second at 1080p.
The state of play in November 2026 is uneven. Laser scanning hardware keeps getting faster and more accurate: Leica Geosystems announced the RTC300, RTC500 and RTC700 terrestrial scanners on 9 June 2026, replacing the RTC360 and ScanStation P-series lines. Photogrammetry suites have matured into hybrid tools. Epic Games renamed RealityCapture to RealityScan with version 2.0 in June 2025 and shipped RealityScan 2.2 in June 2026 with AMD GPU support. Splatting is moving fastest: NavVis added splats as a beta third viewing mode in IVION 12 in August 2026, and OpenUSD v26.03 shipped a native splat schema in March 2026.
NeRF, meanwhile, has largely ceded the production conversation to splats. The research line continues (Instant-NGP, Mip-NeRF 360, Zip-NeRF), but Nerfstudio’s own momentum has shifted to its gsplat library, whose latest PyPI release is 1.5.3, while the nerfstudio package itself has not had a PyPI release since 1.1.5 in November 2024.
The thesis of this article is that the four methods are not competitors on one axis. LiDAR and photogrammetry produce a measured record. NeRF and 3DGS produce a perceptual record. A good reality capture digital twin keeps both, registered in one coordinate frame, and never lets the perceptual layer pretend to be the measured one. The rest of this post shows why, with numbers.
NeRF vs Gaussian Splatting vs Photogrammetry vs LiDAR: The Reference Comparison
NeRF vs Gaussian splatting is a choice between an implicit neural field, which is compact but slow to render, and explicit anisotropic Gaussians, which render in real time but are large. Photogrammetry and LiDAR are different in kind: they output measurable meshes and point clouds. For digital twins, use LiDAR or photogrammetry for geometry and splats for visual context.

Figure 1: The four reality-capture pipelines and where each one ends. Image-based methods share a Structure from Motion front end; only the mesh and point-cloud outputs feed measurement, CAD and BIM.
Figure 1 traces the data from a physical asset to its two possible destinations. Photos or video frames enter Structure from Motion, which solves camera poses and a sparse point cloud. From there the path splits three ways: a NeRF trains an MLP, 3DGS optimises Gaussians, and MVS densifies to a cloud and mesh. Laser returns follow a separate path to a registered point cloud in E57 or LAS. That cloud can also seed a splat. The key point is the bottom row: view synthesis serves human eyes, while measurement, CAD and BIM need the mesh or the point cloud.
The table below is the snippet-ready summary. Figures marked “paper” come from the original publications; vendor specifications are cited inline in later sections.
| Dimension | NeRF | 3D Gaussian Splatting | Photogrammetry (SfM + MVS) | LiDAR scanning (TLS / SLAM) |
|---|---|---|---|---|
| Representation | Implicit MLP radiance field | Explicit anisotropic 3D Gaussians with spherical-harmonic colour | Dense point cloud and textured triangle mesh | Point cloud with intensity, often colourised from panoramas |
| How geometry is obtained | Learned density, implicit | Optimised Gaussian centres and covariances, implicit surface | Triangulated from image correspondences | Measured range per laser pulse |
| Metric scale | None unless poses are scaled | None unless poses are scaled | Needs control points or scale bars | Native metric scale |
| Typical geometric accuracy | Not a measurement product | Not a measurement product | About 1 to 3 times ground sample distance (rule of thumb) | Millimetre class for TLS; about 5 mm local for leading SLAM units |
| Training or processing time | Original paper: about 1 to 2 days on one V100 | Paper: about 35 to 45 minutes on an A6000 for 30K iterations | Minutes to hours, scales with image count | Registration in minutes to hours; capture dominates |
| Rendering speed | Original paper: seconds per frame | Paper: 134 FPS on Mip-NeRF 360 scenes, A6000 | Real time once meshed and LOD-ed | Real time with octree streaming |
| Storage per scene | About 5 MB of network weights (paper) | Hundreds of MB raw PLY; about 10x smaller as SPZ | Tens of MB to GB mesh plus textures | GB to TB of points |
| Reflective, thin, transparent surfaces | Good visually | Good visually, floaters near glass | Weak; holes and noise | Weak on glass and mirrors; good on thin pipes |
| Editability | Very poor | Fair; crop, delete, transform clusters | Good; mesh editing tools | Good; segment, classify, clip |
| CAD and BIM compatibility | None directly | None directly; reference layer only | Mesh import; scan-to-BIM reference | Industry standard for scan-to-BIM |
| Interchange formats | Framework checkpoints only | PLY, SPZ, OpenUSD ParticleField, glTF KHR_gaussian_splatting (RC) | OBJ, FBX, glTF, 3D Tiles, LAS | E57, LAS/LAZ, RCP, 3D Tiles |
| Production maturity, Nov 2026 | Research and VFX | Fast-maturing visualisation | Mature | Mature |
The representation decides what you can ask the twin
Every method stores a scene, but each stores a different answer to the question “what is here?”. LiDAR stores where light came back from: a sample of surface positions with a known error model. Photogrammetry stores where image features agree: a surface hypothesis built from triangulation, which is only as good as texture and overlap. A NeRF stores a function that answers “what colour and how opaque would this point look from this direction”, and a splat stores a large set of soft, coloured, oriented blobs whose blended projection reproduces the training photos.
The consequence is practical. A query such as “what is the clearance between this cable tray and that duct?” requires a surface. LiDAR and photogrammetry give you one explicitly. A NeRF gives you a density field from which a surface can be extracted with marching cubes at an arbitrary threshold, and a splat gives you Gaussians whose centres do not lie on the surface in any guaranteed way. You can extract geometry from both. However, the extracted surface is a by-product of an optimisation that was rewarded for matching pixels, not for matching the physical wall.
Image quality and geometric truth are different objectives
The 3DGS paper optimises an L1 plus D-SSIM photometric loss. NeRF optimises a squared error on rendered pixel colour. Neither loss contains a term for metric geometry. A Gaussian can sit 20 mm in front of a wall and still be “correct” if, from every training camera, it produces the right colour. This is not a bug. It is the design, and it is why view-dependent effects such as sheen on stainless steel look so convincing.
Research has narrowed the gap. Variants such as 2D Gaussian Splatting, which flattens Gaussians into oriented discs, and methods that add depth or normal regularisation produce better surfaces. Splats can also be seeded from a LiDAR point cloud rather than SfM points, which anchors them near real surfaces. Still, none of the mainstream production tools publishes a geometric accuracy specification for splat geometry that a surveyor could sign. Until one does, treat splat geometry as unverified for measurement purposes.
Scale and coordinate frame are the hidden third axis
Image-only pipelines recover shape up to an unknown scale, rotation and translation. SfM gives you a consistent but arbitrary frame. To put a photogrammetric model, a NeRF or a splat into a site coordinate system, you need ground control points, scale bars, RTK-tagged camera positions or a registration to a LiDAR cloud. LiDAR measures range in metres, so scale is native, and a total-station-controlled traverse ties it to the site grid.
This matters more for digital twins than for visual effects. A twin is useful when IoT tags, CMMS asset IDs and BIM elements all line up at the same coordinates. If the splat floats in its own frame, the operator sees a beautiful scene where the sensor icon sits half a metre off the valve. Registration is not optional. It is the integration contract for a reality capture digital twin.
Deeper Analysis: Accuracy, Speed, Cost and Formats in Numbers
This section works through the four axes that drive the decision, with numbers you can check. Where a figure is a rule of thumb or an illustration, it is labelled.
Metric accuracy: what each method can sign off
LiDAR is the only method in this comparison that publishes a geometric error specification by default. Leica’s documentation for the 2026 RTC series lists a 3D point accuracy of 1.5 mm at 10 m and 3.8 mm at 50 m, with range accuracy of 1.2 mm plus 10 ppm, across the RTC300, RTC500 and RTC700. Mobile SLAM scanners trade some accuracy for walking speed. NavVis lists a 5 mm local point-cloud accuracy for the VLX 3, which carries two 32-layer lidar sensors at 1.28 million points per second each. Treat these as vendor specifications under stated conditions; real projects add registration error, target error and surface effects.
Photogrammetry accuracy is governed by ground sample distance (GSD), the footprint of one pixel on the object. GSD equals pixel pitch times distance divided by focal length. As an illustration, a 4.4 µm pixel, a 24 mm lens and a 3 m standoff give 4.4e-6 × 3 / 0.024, or 0.55 mm per pixel. A common field rule of thumb is that well-controlled photogrammetry achieves roughly one to three times GSD in object space, so about 0.5 to 1.7 mm in this example. That rule assumes good texture, strong overlap and ground control. It collapses on painted steel, white walls and shiny pipes.
NeRF and 3DGS inherit the photogrammetric camera solve. Their scale and pose accuracy can be no better than the SfM or LiDAR registration underneath them. Their surface accuracy is a further, unpublished error on top. In practice, the honest specification for a splat is “visual accuracy is high; geometric accuracy is unverified unless you test it against a measured reference”. If you need to report it, compute cloud-to-cloud distances between extracted splat centres and a TLS reference on check areas and publish the distribution.
Processing and rendering speed: where the hours go
The original NeRF needed 100,000 to 300,000 iterations, about 1 to 2 days on a single NVIDIA V100, and took seconds per frame to render. Instant-NGP cut training to minutes with multiresolution hash encodings, but rendering still relies on ray marching. The 3DGS paper’s Table 1 shows the change of regime on the Mip-NeRF 360 dataset, all on an A6000 GPU except where the authors copied Mip-NeRF 360’s own numbers.
| Method (Mip-NeRF 360 dataset, Kerbl et al. 2023, Table 1) | PSNR | Training time | Rendering FPS | Memory |
|---|---|---|---|---|
| Plenoxels | 23.08 | 25 min 49 s | 6.79 | 2.1 GB |
| Instant-NGP Base | 25.30 | 5 min 37 s | 11.7 | 13 MB |
| Instant-NGP Big | 25.59 | 7 min 30 s | 9.43 | 48 MB |
| Mip-NeRF 360 | 27.69 | 48 h | 0.06 | 8.6 MB |
| 3DGS, 7K iterations | 25.60 | 6 min 25 s | 160 | 523 MB |
| 3DGS, 30K iterations | 27.21 | 41 min 33 s | 134 | 734 MB |
Methodology caveat: these are research benchmarks on curated scenes of a few hundred photos at reduced resolution, measured in 2023 on one GPU class. The Mip-NeRF 360 48-hour figure is the authors’ equivalent of 12 hours on a 4-GPU A100 node. They are not throughput figures for a 5,000-image factory capture, and tool versions in 2026 are faster.
Read the table for its shape, not its decimals. NeRF-class methods are small and slow to render; 3DGS is large and fast. Instant-NGP shows that NeRFs can train quickly, but sub-12 FPS rendering is the wrong regime for a VR walkthrough, where common headsets refresh at 72 to 90 Hz and expect a frame for each eye every refresh. 3DGS at 134 to 160 FPS on a workstation GPU fits that budget, and the industry has noticed.
For photogrammetry and LiDAR, field time dominates. A TLS station takes minutes; Leica lists roughly 1 minute 40 seconds for a full dome scan with HDR imagery at 6 mm at 10 m on the RTC500 and RTC700. A plant room might need 20 to 60 stations depending on occlusion, so the scan alone is a day of work before registration. Mobile SLAM collapses that to a walk at roughly walking pace, at the cost of the accuracy noted above. Photogrammetry processing scales with image count and resolution and runs from minutes for an object to many hours for a site.
Storage: a worked splat size calculation
Splats are large because every Gaussian carries its own parameters. In the reference implementation, each Gaussian stores a 3D position (3 floats), a scale (3), a rotation quaternion (4), an opacity (1) and spherical-harmonic colour up to degree 3, which is 16 coefficients per RGB channel (48 floats). That is 59 floats. The reference PLY export also writes an unused normal (3 floats), so 62 float32 values, or 248 bytes per Gaussian.
The paper reports 1 to 5 million Gaussians across its test scenes. At 3 million Gaussians, 3,000,000 × 248 bytes is about 744 MB, which matches the 734 MB average in Table 1. Niantic’s open-source SPZ format, released under the MIT licence, quantises positions to 24-bit fixed point, scales to 8-bit log values, rotations to three 10-bit components and spherical harmonics to 4 or 5 bits by default. Its README states files are “typically around 10x smaller” than PLY, so the example shrinks to roughly 75 MB. That is the difference between a splat that streams to a browser on a site network and one that does not.
A NeRF of the same scene would be a few megabytes of weights. However, those megabytes need a GPU and a ray-marching renderer to see anything, and no generic viewer opens them. In a twin platform, file size is only half of the cost; universal readability is the other half.
Cost: a labour-hours model, not a price list
Hardware and licence prices for scanners and software are quote-based and change often, so this article does not quote them. RealityScan’s licence is free for individuals and companies under USD 1 million in annual gross revenue and requires a subscription above that threshold. Postshot, gsplat and Nerfstudio are free to download, with commercial terms that differ by tool; check each licence before production use.
The better cost model is labour. As an illustration for a 1,500 m² mechanical room, assume a two-person TLS crew needs one day on site and half a day of registration. Add half a day of photography for a splat layer, two to four GPU hours of training, and an hour of cleaning. The splat adds perhaps 25 to 35 percent to the capture labour and almost nothing to hardware if a capable camera and workstation GPU already exist. Scan-to-BIM modelling of the same room is typically the largest cost line by far, and splats do not reduce it. These figures are illustrative, not benchmarks.
Formats and standards status in November 2026
Point clouds have mature, vendor-neutral formats. ASTM E57 stores scans with their poses and panoramic images. LAS 1.4, maintained by ASPRS, and its LAZ compressed form dominate airborne and increasingly terrestrial work. Meshes travel as OBJ, FBX, glTF 2.0 or as 3D Tiles for streaming; the 3D Tiles, Cesium and point cloud streaming architecture post covers how those stream at city scale.
Splats were format orphans until recently. The de facto format is the reference PLY with custom vertex properties, plus compressed variants such as SPZ. In 2026 two standards moved. The Alliance for OpenUSD’s v26.03 release, announced in March 2026, added the UsdVolParticleField schema family, including UsdVolParticleField3DGaussianSplat, a reference renderer called hdParticleField for usdview, and a py3dgsPlyToUsd.py converter. Khronos announced a release candidate of the KHR_gaussian_splatting glTF 2.0 extension on 3 February 2026, with a companion extension, KHR_gaussian_splatting_compression_spz, developed alongside Esri, Cesium and Niantic Spatial. As of this writing, Khronos’s news feed does not show KHR_gaussian_splatting as ratified; treat it as a release candidate.

Figure 2: A two-question decision flow. Survey tolerance routes you to LiDAR or controlled photogrammetry; the need for real-time viewing routes you to splats; most real twins end in a hybrid.
Figure 2 compresses the decision into two questions. If the deliverable must carry a survey tolerance, indoor plant and MEP-dense spaces go to terrestrial or SLAM LiDAR and open sites go to drone photogrammetry with ground control. If it only needs to be seen, real-time browser or VR viewing goes to 3DGS and offline renders can use the NeRF family. The three production branches converge on a hybrid twin, where the splat is a skin on LiDAR bones.
Decision matrix by digital twin use case
The matrix below scores fitness for five common use cases. “Primary” means it can be the deliverable on its own; “supporting” means it adds value as a layer; “avoid” means it misleads or wastes effort.
| Use case | NeRF | 3D Gaussian Splatting | Photogrammetry | LiDAR scanning | Recommended stack |
|---|---|---|---|---|---|
| As-built verification against design | Avoid | Supporting, visual context only | Primary for exteriors with control | Primary | TLS cloud, deviation map against CAD or IFC |
| Remote inspection and walkthrough | Supporting | Primary | Supporting | Supporting | Splat on LiDAR-registered poses, measure on cloud |
| VR walkthrough and training | Avoid, too slow | Primary | Supporting, meshed | Avoid as sole layer | LOD-streamed splat, collision from simplified mesh |
| Scan-to-BIM | Avoid | Supporting, for modeller context | Supporting | Primary | E57 to Revit or IFC authoring, splat as reference |
| Change detection over time | Avoid | Supporting, visual diff | Primary outdoors | Primary indoors | Cloud-to-cloud distance on registered epochs |
The pattern is consistent. Splats win wherever a human must understand a space quickly. LiDAR wins wherever a number must be defended. Photogrammetry wins outdoors at scale and on textured exteriors. NeRF, in production twin work, is now mostly a research ancestor and a niche tool for offline renders.
The hybrid pipeline: splat skin on LiDAR bones
The most useful architecture in 2026 captures LiDAR and images in one campaign, solves everything in the survey frame, and publishes two layers. The 3D Gaussian splatting reality capture pipeline post goes deep on training settings. Here the focus is how the layers stay aligned.

Figure 3: A hybrid capture sequence. LiDAR and images are registered in one frame, the splat is trained on survey-frame poses, and the twin measures on the cloud while displaying the splat.
Figure 3 shows the sequence. The field crew sets scanner stations and targets, then captures overlapping images. Raw scans go to registration, where cloud-to-cloud alignment and control checks run. Images are posed in the same solve, so camera poses and seed points arrive at the splat trainer already in the survey frame. The trainer optimises Gaussians, typically to around 30,000 iterations. The twin platform receives an E57 or LAS backbone and an SPZ or glTF splat layer, and it routes every measurement to the cloud while rendering the splat.
This is the pattern NavVis adopted in IVION 12, where the splat view supports Mark and Measure “backed by the underlying point cloud”. It is also the pattern OpenUSD Core Spec 1.1 and the CAD-to-twin decisions it forces makes practical: a USD stage can hold the CAD-derived geometry, a point-cloud reference and a ParticleField splat prim as sibling layers in one coordinate system.
Editability and lifecycle: the twin must change when the plant changes
A twin is not a photograph; it must follow the plant through modifications. Meshes and point clouds are easy to edit with mature tools. You can clip a region, replace it with a new scan epoch, classify points and attach metadata. LiDAR epochs also diff cleanly: register two scans and compute cloud-to-cloud distances to flag moved equipment.
Splats are editable in a coarser way. You can crop, delete and transform Gaussian clusters, and tools increasingly support selection by region. Patching a replaced pump into an existing splat is harder, because lighting and exposure baked into spherical-harmonic colour will not match across capture days. In practice teams recapture and retrain a zone, then stitch at a boundary. NeRFs are worse: changing one object means retraining the network, and there is no meaningful partial edit. For lifecycle twins, that pushes NeRF out and makes zone-based splat management a design requirement.
Trade-offs, Gotchas, and What Goes Wrong
The failure modes differ by method, and most are predictable. Plan for them in the capture specification rather than discovering them in the viewer.

Figure 4: Six failure modes seen in reality-capture twins and the mitigation for each.
Figure 4 pairs each failure with its fix. Floaters, the semi-transparent fog that hangs in mid-air in splats and NeRFs, come from regions seen by too few cameras. Mitigate with denser coverage, masking and opacity-based pruning. Scale drift in image-only SfM is fixed with ground control or a LiDAR seed. Glass and mirrors produce view-dependent colour that looks right but encodes no geometry; flag such zones as unmeasurable. Dark, wet or highly absorbent surfaces return weak lidar signals, so add photo coverage or move scanner stations closer. Moving people and forklifts leave ghosts in every method, so capture off shift or use semantic masks.
The sixth failure is organisational. Someone opens the splat, measures a distance with a viewer tool that snaps to Gaussian centres, and puts that number in a work order. The splat looked authoritative, so nobody checked. The fix is governance: label the splat layer as visual-only in the twin, route measurement tools to the point cloud or mesh, and record which layer every reported dimension came from. The ASME V&V 40 credibility guidance for digital twins is a useful frame for stating what fidelity each layer supports.
LiDAR has its own traps. Mixed pixels at depth edges create spurious points between foreground and background. Highly reflective targets saturate, and retro-reflective tape can produce range bias. Mobile SLAM can drift on long corridors with few features; loop closures and control points bound that drift, and the LIO-SAM vs FAST-LIO2 vs KISS-ICP comparison explains why the odometry back end matters.
Photogrammetry fails on low texture. Painted ducts, white walls and stainless tanks give the matcher nothing to correlate, which produces holes or bumpy surfaces. Lighting changes over a capture day shift colours and weaken matches. Lens calibration errors propagate directly into scale. Splats trained on such captures inherit every one of these problems in their poses.
Finally, a cost trap. A splat shrinks the “can you show me the site?” request from a site visit to a link. That visibility creates demand. Expect stakeholders to request splats for every area and every quarter. Budget the recapture cadence and storage before you promise it, and keep the 10x SPZ compression in mind when sizing a server.
Practical Recommendations
Start from the deliverable, not the technology. Write down which decisions the twin must support and which of them need a defensible number. Every decision that needs a number must trace back to LiDAR or controlled photogrammetry. Every decision that needs a person to understand a space can use a splat. Very few industrial decisions need a NeRF in 2026, so treat NeRF as a research option, not a procurement line.
Next, specify one coordinate system for everything. Choose the site grid, set control, and require that images, scans and splats all arrive in it. Make that a contract clause for capture vendors, with check-point residuals reported. Then pick formats with an exit path: E57 or LAS for clouds, glTF or 3D Tiles for meshes, and PLY or SPZ for splats today, with OpenUSD ParticleField or the glTF KHR_gaussian_splatting extension as targets once your viewers support them. If your stack lives in Omniverse, Unreal or Unity, check splat import paths for each; the Omniverse vs Unreal vs Unity digital twin comparison covers the engine side.
Finally, plan the BIM link explicitly. Scan-to-BIM authoring consumes point clouds, not splats; the IFC 4.3 openBIM digital twin architecture explains where the resulting model joins IoT data.
A short checklist for your next capture:
- Define tolerance per deliverable and which layer carries it.
- Capture LiDAR and images in the same campaign with shared control.
- Train splats on survey-frame poses, seeded from the cloud where possible.
- Validate splat geometry on check areas with cloud-to-cloud statistics.
- Publish splats as SPZ or USD ParticleField with LOD; keep PLY as the master.
- Route all measurement tools to the point cloud or mesh layer.
- Version every layer by capture date and zone.
Frequently Asked Questions
Is Gaussian splatting better than NeRF?
For digital twin viewing, yes in most practical cases. 3D Gaussian Splatting renders in real time, reported at 134 FPS on the Mip-NeRF 360 scenes in the original paper, against seconds per frame for the original NeRF. It also trains in under an hour on a workstation GPU. NeRF is far smaller, a few megabytes of weights against hundreds of megabytes of Gaussians, and remains useful in research and offline rendering. Neither method gives you surveyed geometry, so the choice between them is about viewing performance and tooling, not accuracy.
Is Gaussian splatting accurate enough for measurements?
Not by default. Splats are optimised to reproduce photographs, so a Gaussian can sit some distance off a real surface and still render correctly. No mainstream tool publishes a survey-grade accuracy specification for splat geometry as of November 2026. The safe pattern is to train splats on poses registered to a LiDAR cloud or controlled photogrammetry, and to measure on that cloud. If you must measure on splats, validate them against a terrestrial laser scan on check areas and report the distance distribution alongside every number.
LiDAR vs photogrammetry: which is more accurate?
Terrestrial LiDAR is usually more accurate indoors and on low-texture surfaces, because it measures range directly. Leica lists 1.5 mm 3D point accuracy at 10 m for its 2026 RTC series. Photogrammetry can match or beat that on well-textured objects at close range with good control, since accuracy scales with ground sample distance. Outdoors at site scale, drone photogrammetry is often the more economical choice. Indoors among pipes, ducts and painted steel, LiDAR is the more reliable backbone, and photogrammetry becomes a texture and splat source.
Can I convert a Gaussian splat into a BIM or CAD model?
Not directly. BIM and CAD tools expect surfaces, solids and classified elements, while a splat is a cloud of soft, coloured Gaussians. You can extract a mesh or points from a splat, but the result is noisy and its accuracy is unverified. The productive route is scan-to-BIM from an E57 or LAS point cloud, with the splat loaded alongside as a visual reference for modellers. OpenUSD can hold both layers in one stage, which helps review, but it does not turn Gaussians into IFC elements.
What file formats support Gaussian splats in 2026?
The working formats are the reference PLY layout and compressed variants such as Niantic’s MIT-licensed SPZ, which is typically about 10 times smaller than PLY. OpenUSD v26.03 added a native UsdVolParticleField3DGaussianSplat schema in March 2026. Khronos published the KHR_gaussian_splatting glTF 2.0 extension as a release candidate in February 2026, with an SPZ compression companion. Check which of these your viewer, engine and twin platform actually read before you standardise, since support still varies by product.
Do I still need a laser scanner if I can make splats from a phone?
For a visual record, a phone or camera capture can be enough. For a digital twin that drives maintenance, clash checks or engineering changes, you still need a measured backbone. Phone captures lack metric scale unless you add control, and they drift over large areas. Many teams now use a hybrid: a mobile or terrestrial scanner for geometry plus a camera pass for the splat. Some scanners, including NavVis devices with IVION 12, are starting to produce both layers from one workflow.
Further Reading
- Our end-to-end 3D Gaussian splatting reality capture pipeline for digital twins covers capture settings, training and publishing in depth.
- How lidar actually works: the physics behind range, noise and reflectivity explains the error sources behind every point cloud.
- OpenUSD Core Spec 1.1 and v26.08: CAD-to-digital-twin decisions shows how USD stages compose CAD, clouds and splats.
- 3MF vs STEP AP242 vs glTF vs JT for lightweight CAD in PLM compares the mesh-side formats.
- External: Kerbl et al., 3D Gaussian Splatting for Real-Time Radiance Field Rendering and the Alliance for OpenUSD v26.03 release notes.
By Riju — about
