Clients rarely ask for a “point cloud,” a “mesh,” and a “gaussian splat” as three cleanly named products. They ask for a 3D model, a digital twin, something they can spin, something they can measure, or “whatever you usually deliver.” Processing menus list all three next to each other. Wish lists treat them as interchangeable. They are not.
Gaussian splat vs mesh vs point cloud is a comparison of representations. Same photos can feed more than one of them. Same flight is not the same deliverable. If you already need the one-sentence definition of a splat, start with What Is a Gaussian Splat? (And What It Is Not). This post is the head-to-head: what each object stores, when each is the right ask, and what clients confuse when the scope says “3D.”
Pilot Ledger is ops and CRM for commercial pilots — quoting, jobs, process handoffs — not a rendering engine. The field job here is naming the product before you fly and before you price.
Three representations, not three skins on one file
Strip the marketing labels. Ask what sits in the file.
A point cloud is discrete samples with coordinates (and often color, intensity, or classification). Each point is a measured or reconstructed sample in space. Dense photogrammetry clouds and lidar clouds still speak the language of many survey, inspection, and volume workflows. The object is points — not a continuous skin, and not a photoreal orbit.
A mesh is a surface. Triangles (or similar faces) approximate the ground, a stockpile, a facade, or a roof. When that mesh was flown and controlled for mapping, you can often measure along it, drape texture, export something CAD or GIS expects, and defend numbers against check shots. The object is a surface — geometry you can treat as continuous for many commercial tasks.
A gaussian splat is a view-centric appearance model — soft 3D blobs (radiance-field style primitives) optimized so a dedicated viewer looks photo-real from the camera poses you actually flew. Position, covariance, opacity, and often view-dependent color make the orbit look right near the path. The object is not a triangle surface and not a LAS you classify by default. Meshes and splats are different representations even when they share the same imagery.
That teaching line matters for scopes: a splat is not automatically a quantitative survey surface. Looks right in a viewer is not the same as a controlled surface or a trusted point product.
What clients confuse
Confusion almost always starts with shared words.
“3D model.” To a superintendent that can mean a phone-friendly orbit. To a CAD tech it means a surface they can click. To a surveyor it may mean points or a mesh tied to control. One phrase, three products. If the quote uses “3D model” as a single line item, you already inherited the argument.
“Digital twin” and “as-built.” Those are project goals, not file formats. An as-built conversation can need plan-view ortho, a dense cloud, a controlled mesh, a walkthrough splat — or a package. Naming the twin does not name the deliverable.
“Looks real, so it must be measurable.” Photoreal foliage in a splat viewer does not authorize stakeout coordinates or stockpile quantities. A textured mesh can look “software-3D” without looking like a photo, and still be the right measurable product. A dense cloud can look sparse and still be the right input for classification or volume.
Same photos = same product. Shared imagery is a processing convenience, not a scope substitute. If they need a walkthrough and a volume, that is two asks — even when one sortie feeds both.
When each is the right ask
Point cloud when the workflow lives in samples: classification, breaklines, LAS/LAZ handoff, volume or change detection specified from points, or lidar/dense photogrammetry named in the contract. Points answer where you sampled and what attributes ride with those samples. They are not a walkthrough by themselves, and not a closed CAD surface until someone meshed them on purpose.
Mesh when the workflow lives on a surface: cut-fill, stockpile surface, grade check, facade quantities, sheets that expect a continuous model, or export into CAD/GIS/BIM that wants a mesh. Independent check shots and a control story usually sit in the money conversation. A pretty texture does not turn a mesh into a splat. A photoreal splat does not turn soft blobs into survey geometry.
Gaussian splat when the client is buying look and walkthrough on purpose: stakeholders who will never open CAD, indoor or site orbits that matter more than stakeout from that file, or a scope that says visualization / “looks like the site.” Primary value is appearance near the flight path; off-path invention stays invention. Do not extract stakeout from a splat viewer. Do not close a survey-scope job with a pretty orbit alone.
What fails when you treat one like another
Splat treated as mesh or survey surface. The meeting looks great. Then someone asks for a quantity, a setback, or a stakeout point from the viewer. You either refuse and look oversold, or you hand out invented coordinates and own the liability.
Mesh treated as splat. The surface may be honest. The client wanted a photoreal orbit for executives. Textured mesh often reads as “3D software,” not “I am standing on the site.” Wrong product for the buyer even when geometry is stronger.
Point cloud treated as either. A dense cloud is not a walkthrough product by itself. It is also not a finished CAD surface until meshed or surfaced for that workflow. Shipping LAS and calling either job done is the same category error with a different extension.
Vague capture for all three. Mapping grids optimized for ortho and DSM are not automatically the capture you want for a photoreal splat. Indoor walkthrough capture is not automatically a survey mesh flight. If you promise more than one product from one pass, write which is primary and which is bonus.
GNSS discipline still applies when placement matters. RTK, PPK, and PPP / OPUS decide whether photo positions and control belong on someone else’s frame. A splat does not forgive a float-heavy flight any more than a mesh or cloud does.
Ortho still sits beside the three
Plan-view orthomosaic remains the map-like image civil sets and overlays expect when georeferencing is done right. Free-orbit 3D — mesh, cloud, or splat — does not replace that sheet language. Many jobs need ortho plus one 3D representation. Say both in the scope.
Scope language that survives the change order
Rewrite fuzzy asks before you fly:
- “3D model” → splat walkthrough, controlled mesh, dense/lidar point cloud, or a named package.
- “Digital twin / as-built” → which files, which viewer, which control story, which checks.
- “Measurable” → what will be measured, by whom, on which product, against which check shots.
- “Looks like the site” → splat (or textured visualization) as appearance — not a substitute for survey products.
Before you quote: What did the client name? Who opens the file next? Are check shots part of the money conversation? Are you pricing one product or two from related imagery? Can you say in one sentence which product is primary?
If you cannot answer those without hedging, you are not ready to pick gaussian splat vs mesh vs point cloud. Fix the scope before you fix the processing settings.
When the representation question is settled, the remaining work is operational: quote the named deliverables, track the job through processing and handoff, and keep the client conversation tied to the products you sold. That ops layer is what Pilot Ledger is built for — CRM, quoting, and job workflow for commercial pilots — so the file-type argument does not become a payment argument later. For product follow-ups, Ask Mav on pilotledger.com.