The Future of As-Built Assets: Unitec student resource

A building is a physical asset and a dataset. The data decides how well the building can be designed, altered, maintained and remembered. The better the data, the longer and more usefully the building can be understood.

• Design intent. Drawings and models describe what was planned.
• As-built reality. Sites change during construction, and again every year of use.
• Handover. Information arrives as folders of files, not as data linked to the asset.

Spatial data closes the gap by recording what is actually there.

• Drone photogrammetry. Photos into 3D models. Best for roofs, facades and large sites.
• Laser scanning. Dense, precise point clouds. Best for interiors and fine detail.
• Handheld SLAM. Fast walk-through scans. Best for complex or occupied spaces.
• 360° imagery. Panoramic visual records. Best for progress and condition.
• Survey control. Fixed reference points that tie datasets together.

Different tools, one goal: a measured record exists.

A point cloud is a large set of measured points, each with a position in 3D space and often a colour. Laser scanners produce them directly, and photo processing can produce them too. They are measurable and accurate, but the files are heavy and hard to read without software. The first video above is a short introduction.

Survey control means fixed reference points, measured with GNSS or a total station, that tie every dataset to one coordinate system. It is the part people forget. Without it, a drone model, a laser scan and a 360 capture of the same building will not line up.

• 01 Capture. Drone, scanner and 360 data collected on site with survey control.
• 02 Process. Images and scans aligned into point clouds, meshes and orthomosaics.
• 03 Interpret. Features classified and modelled: walls, services, defects, scan-to-BIM.
• 04 Deliver. Published to a common data environment as an asset information model.

Data quality is set at step one. Accuracy, coverage and control decide what every later step can trust. Nothing downstream can fix a poor capture.

Scan-to-BIM is the interpretation step: a person or an AI model turns captured geometry into modelled objects such as walls, services and defects. The result carries asset data that a raw point cloud cannot.

• Point cloud. Measurable, accurate geometry. Heavy files, hard to read.
• Textured mesh. Easy to view and share. Surfaces only, no object data.
• BIM model. Objects carrying asset data. Built from capture.
• Gaussian splat. Photoreal and quick to capture. Visual-grade accuracy only.

Choose by the question you need answered, not by what looks best. To measure a clearance, use the point cloud. To show a client, a mesh or a splat travels well. To schedule maintenance, you need a model with objects and data.

Gaussian splatting turns photos or video into a photoreal 3D scene. Instead of building surfaces from triangles or recording points, the scene is stored as millions of small, soft, semi-transparent blobs, each with a position, size, orientation, colour and opacity. They are optimised until the rendered scene matches the source photos from every angle.

The technique comes from research published in 2023 (Kerbl, Kopanas, Leimkühler and Drettakis, "3D Gaussian Splatting for Real-Time Radiance Field Rendering") and is now moving into mainstream software. The second video above is a short introduction.

It captures how a place looks, including the things meshes struggle with: reflections, glass, foliage, weathered texture and fine detail. Capture is quick and needs only a phone, a 360 camera or a drone, so far more people can document far more sites. The result is a walk-through scene that people can understand without training.

• Site and context. A realistic record of the existing conditions around a design before work starts.
• Heritage and conservation. A visual record of a place before restoration, alteration or loss.
• Construction progress. Documenting how a site changes over time.
• Client and community presentation. Walk-through scenes that people can explore for themselves.
• Studio and teaching. Students can build a 3D presentation of a place without a specialist modelling pipeline.

Not reliably. A splat is visual-grade, not survey-grade. It is built to look right, not to hold dimensions. Measurements still come from laser scans and surveys. A splat can be scaled and aligned to survey control, but its accuracy must be checked and stated before anyone relies on it. Do not let a beautiful render stand in for a survey.

Photogrammetry solves for geometry and produces textured surfaces. A splat optimises for appearance and has no clean surfaces at all. That makes splats better at reflective, translucent and fine detail, and worse for anything that needs geometry: clash checks, section cuts, take-offs or building a BIM model.

Not as modelled objects. A splat has no object data and no clean surfaces, so it works as a visual reference beside a model, not as the model. Unified viewers are beginning to show scans, splats, imagery and BIM models in one environment, and the tools listed in the next question already read splats directly.

Support is strongest in geospatial tools, partial in rendering tools, and mostly through plugins in game engines. Status as of October 2026; this changes quickly, so check each vendor.

• Cesium. CesiumJS and Cesium for Unreal added support in March 2026, using the new glTF splat extension.
• Bentley. Splats are available in its cloud services, the iTwin Capture desktop client and Cesium ion.
• Esri ArcGIS. ArcGIS Pro and the ArcGIS Maps SDK for JavaScript have Gaussian splat layers.
• Autodesk Arnold. Native splat rendering since version 7.5.2. Autodesk also helped write the standard, but the sources show no splat support yet in design tools such as Revit.
• Unreal Engine. No native support. Plugins fill the gap, including Cesium for Unreal.
• Unity. No native support. The open-source UnityGaussianSplatting plugin is widely used.
• Blender. Native import and rendering are in development. Add-ons such as the one from KIRI Engine work today.
• NVIDIA Omniverse. Splat support through NVIDIA's NuRec library and OpenUSD.

The glue is a standard: KHR_gaussian_splatting, a glTF extension from Khronos with contributions from Autodesk, Bentley, Esri, NVIDIA and Niantic Spatial. One outlet reports it was ratified in September 2026; check the Khronos glTF registry for its current status.

• Visual-grade accuracy only.
• Views far from the captured camera positions can show blur or stray artefacts.
• Files can be large and editing tools are still immature.
• No object data: a splat cannot tell you what a wall is made of or when it was last serviced.
• The result depends on the capture: coverage, overlap and lighting all show up in the scene.

You can make a usable splat with a phone and no special equipment. The capture matters far more than which app you pick, so practise the technique first.

Before you start

• Pick a still subject. People, pets, cars and moving leaves show up as ghosts or fog.
• Use soft, even light. Overcast sky or shade is ideal. Hard sun shadows and mixed lighting cause patchy results. Keep the light the same for the whole capture.
• Avoid problem surfaces. Mirrors, glass, shiny metal, thin see-through material and plain white walls capture poorly.
• Clean the lens, charge the phone and lock focus and exposure before you begin, if your app allows it.

How to move

• Move slowly. Walking pace or slower. Fast movement causes blur, and blur cannot be fixed later.
• Overlap every shot. Aim for at least 60% overlap between neighbouring frames, so every part of the subject appears in several views.
• Circle the subject. Walk a full orbit at eye level, then another with the phone tilted slightly down, and a third from higher or lower if you can. For a room, walk the perimeter first, then once through the middle.
• Cover the details you care about. Come closer for texture and carvings, step back for context. Hold the phone steady and keep the subject in frame.
• Plan the whole loop, because gaps in coverage become holes in the result. Typical captures run a few minutes and a few hundred frames.

Free software for each step

• Capture and process on the phone, free: Scaniverse (iPhone and Android). It records and processes the splat on the phone itself, in about a minute or two, with no PC. It exports in the open SPZ format. A LiDAR iPhone (12 Pro or later) helps on plain or dark surfaces.
• Cloud apps with limited free tiers: Polycam, KIRI Engine and Luma AI. You upload the capture and they train the splat. Free limits and which features are paid change often, and reviews disagree, so check each app's current plan before you rely on it.
• Higher quality on a computer, free: use the phone only to record the video or photos, then train on a PC. RealityScan (Windows, free for students and small organisations) works out the camera positions, and a trainer such as Brush, LichtFeld Studio or Postshot builds the splat. See the "Free software to process a splat" section below for terms and limits.
• Clean, crop and share, free: SuperSplat. Open the exported PLY (or SPZ) in the browser, remove stray floaters, crop to the subject, then export or publish. See the editing section below.

A simple first project

• Choose a small, still subject outdoors on an overcast day, such as a doorway, a statue or a garden bench.
• Capture it in Scaniverse with two slow orbits, then open the result on the phone to see what is missing.
• Repeat once, slower and with more overlap. Compare the two. Then export it and clean it up in SuperSplat.

Remember the limit covered above: a phone splat is visual-grade, so use it to show and explain a place, not to take measurements. For heritage places, only capture what you have permission to record, and for Māori places wait for an invitation from mana whenua.

Processor (alignment and 3D models)

Professional photogrammetry software, formerly RealityCapture. It aligns photos or scans into camera positions, point clouds and textured 3D models, and can export camera data for splat trainers such as LichtFeld Studio, Postshot and Brush. All features are included in the free licence, which covers students, educators and individuals or companies with under USD 1 million in annual gross revenue. Educators and schools have no revenue limit. Above the threshold a paid seat is required. Windows desktop; free licences come through the official site.

realityscan.com · RealityScan 2.0 coverage

Processor, also edits

A desktop workstation that trains splats from COLMAP datasets and also lets you inspect, select, transform and export them. Open source under GPLv3, so building it from source costs nothing on Windows or Linux. The ready-made Windows download comes through the project's portal for a minimum contribution of $30, which funds development. It needs an NVIDIA GPU (RTX 20-series or newer) and a recent driver.

lichtfeld.io · source on GitHub

Processor

A friendly Windows desktop trainer that imports camera data from COLMAP and RealityScan. The Free plan has no limit on images, video length or model size, and trains and renders up to 4K, but it is non-commercial and saved files only work within the Free plan. Exporting the splat itself (PLY, SPZ) and commercial use need the paid Indie plan. Requires an NVIDIA GPU.

jawset.com · current plans

Processor

A research-grade framework whose splatfacto method trains Gaussian splats. Fully open source and the most flexible option, but it is run from the command line, needs an NVIDIA GPU with CUDA, and setup can be fiddly. It starts from COLMAP geometry when available, which gives better results than random starting points.

Nerfstudio splatfacto docs

Processor, also views

A trainer and viewer built on WebGPU rather than CUDA. It runs on Windows, macOS, Linux, Android and in a browser, on AMD, Nvidia and Intel graphics. Licensed Apache-2.0. It reads COLMAP and Nerfstudio-format datasets and can open PLY splats. Still a young project, so expect rougher edges than the polished desktop tools.

Brush on GitHub

Capture and processor

A phone app from Niantic Spatial that captures and processes splats on the device, with no PC needed. It exports in the open SPZ format. Independent reviews describe the basic app as free; plans and enterprise features may change, so check the current terms. Quality is lower than a desktop trainer, but it is the fastest way to try the technique.

scaniverse.com

Processor (cloud)

Upload photos or video and the service trains the splat for you. Both have free tiers with limits, and reviewers disagree on exactly what is free, so read the current plans before relying on either.

Luma AI · KIRI Engine

Prerequisite step

Free, open-source structure-from-motion software that works out where every photo was taken. Many trainers, including LichtFeld Studio and Brush, start from COLMAP output. It does not make a splat by itself.

colmap.github.io

Editor

PlayCanvas's free editor runs in the browser and installs as a web app. Use it to clean up, crop, colour-adjust and animate a splat, then export or publish it. MIT licensed. Version 3.0 (September 2026) is rebuilt on WebGPU and needs a current Chrome, Edge, Safari 26 or later, or Firefox with WebGPU enabled.

SuperSplat · open the editor

Editor (converter)

PlayCanvas's open-source tool for scripted, repeatable conversions and batch jobs between splat formats such as PLY, compressed PLY, SOG and a standalone HTML viewer. Run it with npx @playcanvas/splat-transform. Useful when a splat is too large to share.

SplatTransform manual

Processor and editor

Besides training, it supports selecting and transforming Gaussians with undo and redo, and exports to PLY, SOG, SPZ or a standalone HTML viewer. Pricing and details are in the processing section above.

lichtfeld.io

Editor (general 3D)

Blender is free and open source. Native splat import and rendering are in development; add-ons such as the one from KIRI Engine let you bring a splat into a Blender scene and render it today. Good for combining a splat with modelled geometry.

blender.org · KIRI Engine

The main news site for Gaussian splatting and radiance fields, with a directory of platforms and a monthly roundup.

radiancefields.com · monthly newsletter

The announcement of the open standard for storing splats in glTF, backed by Autodesk, Bentley, Esri, NVIDIA and others.

Khronos press release · CG Channel coverage

Why an open standard for splats matters for infrastructure, and how Cesium brought the standard to CesiumJS and Unreal Engine.

Bentley blog · Cesium announcement on Radiance Fields

Release notes for SuperSplat and SplatTransform.

blog.playcanvas.com

Two of the industry sources behind the talk, covering reality capture and the built environment.

AEC Magazine · Cintoo

"3D Gaussian Splatting for Real-Time Radiance Field Rendering" (Kerbl, Kopanas, Leimkühler and Drettakis, 2023) is where the technique began.

Project page

Geospatial World · Watch on YouTube

Bad Decisions Studio · Watch on YouTube

Autodesk Developer · Watch on YouTube

A twin is only as good as its data. Placeholders and look-alike components make a twin unreliable. A twin should mirror what is actually built, and be updated as the asset changes.

The project information model (PIM) is built during design and construction. At handover it becomes the asset information model (AIM), used in operation. This is where your model stops being a design deliverable and starts being someone's operating record.

ISO 19650 is the international standard for managing information across an asset's life. Draft updates to parts 1 to 3 went to consultation in March 2026, with stronger alignment to digital twins, AI workflows and sustainability reporting.

Structured handover means asset data, such as COBie, linked to rooms and equipment, not loose drawings in a folder. The owner can then search by asset or location instead of hunting through documents.

• Capture before change. A measured record should exist before restoration, alteration or loss.
• Use open formats. E57, LAS and IFC outlive any single piece of software.
• Capture again over time. Repeat surveys show movement, wear and unrecorded changes.
• Share with custodians. Owners, communities and schools can all use the same record.

Clark House (Ngaroma) is a heritage home shown as an explorable 3D model on the Tohu Heritage site. Heritage and ageing buildings change slowly and then suddenly: a restoration, a fire, a demolition. A good record captured before that moment is permanent.

CyArk is a non-profit based in Oakland, California, that digitally records, archives and shares the world's cultural heritage. Its stated aim is "to make the world's cultural heritage accessible to new audiences and future generations."

• Founded 2003 by Ben Kacyra, an early LiDAR pioneer and founder of Cyra Technologies. That makes it active for over two decades.
• Origin. Accounts of its history link the founding to the destruction of the Bamiyan Buddhas in 2001, which led Kacyra to apply laser-scanning technology to cultural heritage.
• Reach. CyArk reports working with local partners at more than 200 sites in over 40 countries, including Angkor Wat, Pompeii, Chichen Itza and Mount Rushmore.
• Method. Field teams capture sites with LiDAR and photogrammetry, then turn the data into archives and public experiences.

It is a useful reference for students because it shows heritage documentation at global scale, with the data kept as an open, lasting record.

cyark.org · CyArk mission · Wikipedia

Tapestry is CyArk's own platform for sharing heritage stories. CyArk captures a place, then Tapestry turns it into a guided 3D experience that anyone can explore in a web browser or on a phone, with nothing to install. The experience combines a photorealistic 3D view with scripted narration, archival images, audio, video and the voices of the people connected to the place.

• Relationship. CyArk is the organisation that documents and archives sites. Tapestry is the public-facing platform where that documentation becomes a story people can visit.
• Launched 2022, according to a National Trust for Historic Preservation article, as a way to tell place-based stories on web and mobile.
• For partners. A site, museum or community can have an experience produced with CyArk and embed it on their own website with a single link.
• Now using splats. In 2026 CyArk published new experiences, including Pompeii and Civita di Bagnoregio and an America 250 collection, that use Gaussian splatting.

Compare this with the split in this resource: measured archive data on one side, and the story that makes people care on the other.

Tapestry · browse the library · how it works · National Trust article · 2026 splat experiences

A measured model tells you where the walls are. It does not tell you why the building exists. Places carry stories: who built them, who lived there, and what they meant. Without the story, a scan is only geometry. People look after what they care about.

In Aotearoa, culture also sets what matters. Communities, iwi and hapū decide what is significant about their places. Māori heritage places are only documented when mana whenua invite it, and they keep ownership of what is made.

• Architects. Accurate existing conditions before design starts, context models for retrofit and adaptive reuse, and evidence for heritage and consent documentation.
• Engineers. Check clearances and deviations from design, fabricate and fit to what was actually built, and compare surveys over time to track movement.
• Asset owners. One spatial record they own, faster handover with data linked to assets, and one source for planning, insurance and maintenance.

The same capture serves all three, which is why paying for it once makes sense.

• 1. How accurate is it, and how was that checked?
• 2. When was it captured, and what changed since?
• 3. Which coordinate system and survey control does it use?
• 4. Who owns it, and will the format still open in ten years?

Accuracy and date matter most for design decisions. The coordinate system matters when datasets must line up. Ownership and format decide how long the record survives.

• Go beyond the brief. Deliver what you were hired for, then guide the owner on what to do with it.
• Explain the data. What was captured, how accurate it is, and what questions it can and cannot answer.
• Set up for the long term. Storage, open formats, update cycles, and who keeps the record current.
• Bring everyone along. Everyone who works on the asset, directly or indirectly, shapes how well it is managed.

Own the problem end to end. It raises the chance the asset is managed in the best way possible.