Lesson 4.4Lesson 4.4 · Neural Capture: NeRF & Gaussian Splatting
Where Neural Capture Fits
The honest placement: neural capture is a superb tool for seeing, presenting and experiencing a space, a poor tool for measuring it, and a brilliant complement to the measured capture it cannot yet replace
Neural capture is not a better scanner and not a worse one — it is a different instrument. The skill is knowing which jobs it was built for, which jobs it will quietly ruin, and how to run it alongside the tools it cannot replace.
Across this module we have built two neural techniques — NeRF and Gaussian splatting — and reached the same verdict about both: astonishing photorealism, real limits on measurement. This final lesson turns that verdict into practical judgement. The question is not "is neural capture good?" — it is *good at what?* — and the honest answer places it precisely.
Neural capture is outstanding for a whole class of work: visualisation, presentation, immersive VR and AR, synthesising new viewpoints, and capturing the visual richness of spaces that measured methods render coldly or struggle with entirely. It is not, today, a substitute for laser scanning or photogrammetry wherever metric accuracy is required — coordination, setting-out, fabrication, boundaries, structural work. And the most mature position of all is not to pick a side but to *combine*: let measured capture carry the geometry and neural capture carry the experience. Hold that placement, stay honest about the limits, and watch a field that is changing month by month.
Where neural capture fits: see with it, don't measure with it. Combine layers, label what's measurable, defer binding to a surveyor. Watch the field.
Where it excels: visualisation, presentation, VR/AR, view synthesis
Begin with the bright side, stated precisely, because neural capture genuinely does things the rest of this course's methods cannot. Its home turf is everywhere the goal is to see and experience a space rather than to measure it.
Visualisation and presentation. A NeRF or Gaussian splat gives a client, a jury or a stakeholder a photorealistic, fly-anywhere model of a real place — captured from a phone or camera in hours, not days. For winning a competition, explaining a design in its real context, or showing the before of a renovation, nothing conveys the *feel* of a space as vividly. Where a coloured point cloud looks clinical and a textured mesh looks flat, a splat looks like being there.
VR and AR. Because Gaussian splatting in particular renders in real time, it drops naturally into virtual-reality walkthroughs and augmented-reality overlays, letting people explore a captured space interactively in a headset or a browser. For immersive design review, remote site familiarisation, and client experiences, this is a capability measured capture simply does not offer in the same photoreal, low-cost way.
Novel view synthesis. The defining trick — generating convincing images from viewpoints no camera occupied — is useful in its own right: producing smooth cinematic fly-throughs, filling in views for media and marketing, and creating visual content from a sparse set of photographs.
Capturing visual richness. Neural methods shine on exactly the subjects that defeat scanners: reflective, glossy and glass surfaces; soft and fuzzy things like foliage, fabric and hair; and intricate ornament — carved brackets, a pierced jali screen, a weathered facade — where the *appearance* carries the value. For heritage visual documentation especially, a neural capture records the look and character of a surface with a fidelity that complements the hard geometry of a laser scan. In all these cases the job is fidelity of appearance, and neural capture is, right now, one of the best tools we have for it.
Neural capture's home turf: SEE and EXPERIENCE. Presentation, VR/AR, fly-throughs, reflections, foliage, carved ornament. Appearance is the job.
Where it does not belong: anything that must be measured
Now the boundary, drawn firmly, because this is where misuse does real damage. Wherever the deliverable must carry trustworthy dimensions, neural capture is — today — the wrong primary tool, for the reasons the whole module has built: it is optimised to reproduce appearance, its geometry is approximate, and it has no inherent scale and no honest error figure.
Concretely, do not rely on a NeRF or a splat as the metric basis for: coordination (clash-checking new services against existing structure needs real positions); setting-out (building to a neural model's dimensions invites misfit); fabrication and prefabrication (off-site manufacture depends absolutely on true measurements, as Module 0 stressed); scan-to-BIM where the model must be dimensionally reliable (you model from a controlled laser scan or photogrammetry, not from a splat); boundaries, cadastral and legal survey; structural or deformation assessment; and any georeferenced, survey-grade deliverable. In every one of these, a wrong dimension that looked right is exactly the costly failure this course exists to prevent.
The danger is sharpened by neural capture's very excellence. A clinical point cloud does not tempt anyone to over-trust it; a photorealistic, real-time splat absolutely does. A client flying smoothly through a gorgeous model will assume it is accurate, and a colleague under deadline pressure may be tempted to "just grab" a dimension off it. Part of your professional duty is to *label the instrument honestly* — to state, in the deliverable and in the conversation, that the neural capture is a visualisation of high fidelity and uncertain metric truth, and that any measurement must come from, or be verified against, measured capture.
And the firm professional boundary from Module 0 stands undisturbed. Where a result must be legally or structurally binding — a boundary or cadastral survey, a setting-out, a deformation or structural monitoring survey, a georeferenced survey-grade deliverable — that belongs to a licensed surveyor or geospatial professional, working to recognised standards (including the Survey of India framework), verified equipment specifications and, for aerial capture, the Drone Rules / DGCA regime. Neural capture changes none of that. It adds a superb way to *see*; it removes none of the discipline required to *measure*.
How it complements measured capture — combine, don't choose
The most mature practice does not treat this as neural-versus-measured at all; it runs them together, each doing what it is best at. The model to carry is two layers over one project: a *metric backbone* captured by laser scanning or controlled photogrammetry, and a *visual layer* captured neurally — tied, ideally, to the same control so they describe the same place.
Picture a heritage courtyard. A terrestrial laser scan, on proper control, gives you the trustworthy geometry: accurate dimensions, a point cloud you can scan-to-BIM from, a measurable record for conservation and structural work. A Gaussian splat of the same courtyard, captured alongside, gives you the experience: a real-time, photoreal walkthrough that captures the patina of the stone, the play of light, the carved detail, the reflections in a water channel — everything the point cloud renders coldly. Present with the splat; measure on the scan. Your deliverable is both accurate *and* beautiful, and at every moment you know which layer any given number came from.
This complementarity runs the other way too. Neural capture can *extend* measured capture into places it is weak: photogrammetry and laser scans struggle with glass, gloss, water and fine foliage, and a neural capture can document the *appearance* of exactly those elements to sit alongside the hard geometry. Conversely, measured capture can *discipline* neural capture: a known dimension or survey control can impose and check real-world scale on a NeRF or splat, and an accurate scan can be the reference against which you verify, or correct, neural geometry. The two are genuinely better together.
The practical decision rule is simple, and it is the one your whole module has been building toward: ask what the output is *for*. If it must carry dimensions, lead with measured capture. If it needs to look and feel real, lead with neural capture. If it needs both — as serious projects usually do — run both, over shared control, and keep the layers clearly labelled. That is not a compromise; it is the state of the art, and it is how capture-literate practices already work.
Two layers, one project: measured backbone (scan/photogrammetry) for geometry + neural layer (NeRF/splat) for experience. Measure on one, present with the other.
The fast-moving caveat — and the principle that outlasts it
Every specific limit stated in this module comes with an expiry warning, and honesty requires flagging it plainly. Neural capture is among the fastest-moving areas in all of computer vision. NeRF is from 2020; Gaussian splatting, which largely replaced it for real-time use, is from 2023; and the research that attacks exactly the weaknesses we have named — imposing and checking real-world scale, extracting clean and correct meshes, suppressing floaters and artefacts, tying neural captures rigorously to survey control, even pushing toward measurable accuracy — advances month by month. It would be both arrogant and unhelpful to tell you that today's limitations are permanent. Some of them will fall.
So the right stance is neither breathless nor dismissive. Do not assume the technology cannot improve; equally, do not believe a vendor demo or a slick splat that "it is now survey-grade" without proof. The test that does not expire is evidence: a specific tool earns metric trust for a specific job only when it is demonstrably validated against independent measured data and, where the result is binding, signed off by a licensed surveyor working to recognised standards. Re-check current capabilities before you rely on any neural method for measurement; treat this lesson's caveats as a snapshot as of writing; and let the proof, not the polish, decide.
What *will* outlast every technique is the principle this module exists to teach, and it is worth stating one last time. A representation trained to reproduce appearance is not automatically a measurement. Looking real and being metrically true are different properties, and knowing which one you have — and which one the job needs — is the core judgement of the capture-literate professional. Learn the principles, not just the products: what a representation is, what it optimises for, what it can and cannot be trusted for, and how to verify. Do that, and whatever next year's breakthrough is called, you will be able to place it honestly, use it for what it is good at, and hand the binding measurement, as ever, to the people and instruments whose job that is.
The limits are a snapshot; the principle outlasts them: appearance != measurement. Let evidence, not polish, decide. Re-check every year.
Fit to purpose
Matching method to the deliverable
Neural capture for seeing (visualisation, VR/AR, view synthesis, visual richness); measured capture for measuring. Ask what the output is FOR before choosing.
Complement, not replace
Combining measured and neural capture
Best practice runs a measured metric backbone plus a neural visual layer over shared control; label which layer is measurable. Measured capture per Modules 2-3.
Binding results to professionals
Boundaries, setting-out, structural, georeferencing, aerial
Unchanged by neural methods: defer to a licensed surveyor, verified specs and the governing rules (Survey of India; Drone Rules / DGCA for aerial). Module 9.3-9.4.
Evidence over polish
Fast-moving accuracy claims
A neural tool earns metric trust only when validated against independent measured data; re-verify current capabilities rather than believing demos or treating this snapshot as permanent.
Workshop — write the capture plan for a real project using both layers
This module ends where practice begins: deciding, for a real project, what to capture how. You will write a two-layer capture plan that places neural and measured capture honestly against the project's actual needs.
Your notes from Lessons 4.1-4.3 and a real or realistic project. No capture hardware required; the deliverable is the reasoning.
Goal: a defensible capture plan that uses neural capture for what it is good at and measured capture for the rest Inputs: a real or realistic project (a renovation, heritage job, fit-out or competition) + this lesson + your triage and comparison notes from 4.1-4.3 Time: ~50 minutes
- 1Define the project's deliverables: list everything the capture must ultimately support — client presentation, VR tour, coordination model, scan-to-BIM, fabrication, conservation record, any boundary or structural need.
- 2Split by job: for each deliverable, decide whether it needs measured geometry, photoreal appearance, or both, using the decision rule 'what is the output FOR?'.
- 3Specify the metric backbone: state what you would capture by laser scanning or controlled photogrammetry, to roughly what level of accuracy/detail the job needs, and where it must tie to control (noting this is illustrative, not a spec).
- 4Specify the neural visual layer: state what you would capture with NeRF or Gaussian splatting, for which deliverables, and how you would tie or scale it to the measured backbone if possible.
- 5Write the honesty and hand-off section: state plainly which outputs are measurable and which are visualisation, which dimensions you would verify and how, and exactly where a licensed surveyor (and, for any drone work, the Drone Rules / DGCA) is required.
You’ll walk away with
A one-to-two page two-layer capture plan: deliverables split by job, a measured backbone spec, a neural visual-layer spec, and an explicit honesty-and-hand-off section naming what is measurable, what must be verified, and where a surveyor is required. This is the synthesis of the whole module and a genuine practice artefact.
Three altitudes on the same idea
Read the band that fits you — or all three.
Place neural capture as your experiential layer and measured capture as your metric backbone, and run them together. For presentation, jury and client walkthroughs, VR/AR, design-review context and heritage visual documentation, a NeRF or splat is superb and cheap to capture. For the dimensions that drive coordination, setting-out, fabrication and scan-to-BIM, lead with laser scanning or controlled photogrammetry on proper control, and, ideally, tie the neural capture to that same control so both describe one place. Label every deliverable honestly about which layer is measurable. And defer everything binding — boundaries, setting-out, structural or deformation work, georeferenced survey-grade deliverables, aerial capture under the Drone Rules — to a licensed surveyor and verified specs. Watch the field: re-check neural accuracy claims against evidence before ever relying on them to measure.
Use neural capture for the experience of an interior and measured capture for its dimensions — and you get the best of both. A Gaussian splat gives a client an immersive, photoreal, real-time tour of their space, its materials and light captured vividly; that is its job, and it does it brilliantly for presentation and before/after storytelling. For the true ceiling heights, alcove widths and out-of-square angles that joinery and fit-out depend on, capture measured data (phone LiDAR with checks, a tape, or commissioned photogrammetry) and never read those numbers off the splat. Where both matter, capture both and keep them labelled. Flag anything binding for professional survey, and remember the field is moving fast — verify any claim that a neural tool is now accurate enough to measure from before you trust it.
The capstone skill of this module is placement: knowing that neural capture is for seeing and measured capture is for measuring, and that the best practice combines them. Be able to list where neural capture excels (visualisation, presentation, VR/AR, novel views, visual richness like reflections and ornament) and where it must not lead (coordination, setting-out, fabrication, scan-to-BIM, boundaries, structural work). Explain the two-layer model — metric backbone plus visual layer over shared control — and why it beats choosing a side. And carry the principle that outlasts every tool: a representation trained to reproduce appearance is not automatically a measurement, so let evidence, not polish, decide, and re-check a fast-moving field. This judgement, clearly argued, is exactly what a capture-literate portfolio shows.
“Neural capture has basically won — it is photorealistic, cheap, real-time and improving fast, so the sensible move is to standardise on NeRF or Gaussian splatting and phase out expensive laser scanning and fiddly photogrammetry for capturing buildings.”
Do it yourself
No tools needed — place it honestly.
- 1List four things neural capture excels at, and say what they have in common (hint: seeing, not measuring).
- 2List four jobs where neural capture must not be the metric basis, and say what goes wrong if it is.
- 3Explain the two-layer model — metric backbone plus visual layer — and why combining beats choosing.
- 4Give one way neural capture complements measured capture, and one way measured capture disciplines neural capture.
- 5State the principle that outlasts any specific technique, and how you would decide whether a new neural tool can be trusted to measure.
The one line to carry out
Peer-reviewed journals & authoritative standards
- 01Novel view synthesis — Wikipedia — Novel view synthesis, 2026.
- 02Gaussian splatting — Wikipedia — Gaussian splatting, 2026.
- 03Neural radiance field — Wikipedia — Neural radiance field, 2026.
- 04Surveying — Wikipedia — Surveying, 2026.
- 05Accuracy and precision — Wikipedia — Accuracy and precision, 2026.
That completes the neural turn: NeRF, Gaussian splatting, and an honest placement of both. With the full family of capture methods now in hand — laser, photogrammetric and neural — the course turns to the data they all produce and how to tame it: point clouds and meshes, their registration, cleaning, formats and sheer scale.
The author
Amogh N P
Architect, interior designer, and creative polymath. Studio Matrx began in his notebooks — his vision of design made honest, useful, and open to everyone. Its Academy is written and taught in his memory, and free, forever.
More about Amogh →