Designing Inside the Splat
Any photographer will tell you the best camera is the one you have on you.
Reality capture works the same way. Most days, the only scanner in my pocket is my phone. So when I'm standing in a space and need to capture it now, I open Scaniverse and record. That's the best I'll get out of that moment. It isn't precise, but it beats forty unlabelled photos.
This post is about what happens after that: when a project is worth scanning properly, and you intend to design inside the result. Not beside it. Inside it.
Choosing the scanner
The phone gets you started. As soon as the project justifies it, we switch to a dedicated scanner.
The PortalCam is the one you'll actually carry. It fits in a shoulder bag, a rucksack, even a satchel. But the Lixel K1, K2 and L3 are all more accurate, so when precision matters, the PortalCam stays in the bag.
Large, repetitive spaces change the calculation again. Warehouses. Airport terminals. Factory floors. Anywhere that looks the same in every direction. A short-range scanner loses track of itself there. It sees repeated geometry and repeated visuals, can't tell one bay from the next, and the scan starts to drift. A long-reach scanner like the L3, with a range of up to 120 m, can lock onto features far into the distance and hold its position as you move.
So the question isn't which scanner is best. It's which one you have, and what level of accuracy the project needs.
How we scan
When we know a model will go into the space, we follow the same rules every time:
All the lights on, all the doors open.
Windows closed where possible, so drafts and wind don't disturb the scene.
Scan at the quietest time of day. Agree it with the client and the capture team up front. People walking through a scan leave ghosts.
Plan your route before you start. Scan the building the way you want to navigate it later.
Work in 10-minute segments, so you can check the data is good before you move on.
Our longest continuous scan ran 45 minutes, an indoor-outdoor capture on a deadline that left no time for check-in breaks. It worked. It was also a risk I wouldn't recommend.
A few habits make a surprising difference:
Go through doorways sideways. Hold the scanner closer to your chest and step through side-on, so it sees the room you're leaving and the room you're entering at the same time. This matters most where lighting changes between zones.
Watch your pace. One to two metres a second is fine for large areas. For photographic detail, slow down to 0.8–1.0 m/s. It feels painfully slow. It's worth it.
Walk an S-pattern in front of products. In retail, you want the space and the merchandise. Weaving in front of a fixture gives the scanner far more angles than a straight pass.
On light
Indoors, the rule is simple: every light on. Good general illumination makes for a better splat.
Malls are harder. Mall lighting is rarely good. Some centres run warm fluorescents, some cold, and every store within a store has its own lighting. If you scan one section of a department store, scan its perimeter too, under that section's lighting. Otherwise, the edges of your zone carry a different light profile from its centre, and the splat reads patchy.
Outdoors is a different conversation. South Africa is among the sunniest countries in the world, and outside the rainy months we rarely deal with grey skies. Scan early and you get warm, natural daylight. Not a cloudy, drowsy, depressing splat.
That matters more than it sounds. You're selling a vision of the future, and you want it realistic, but realistic on the client's best possible day. So we recommend a golden-hour window: 7–9 am or 4–6 pm. Warm light, and the brightness doesn't overwhelm.
(If the client's best possible day is rain outside and Netflix in bed, that's a design decision too.)
Getting it into SketchUp
We bring scans in as LCC2, XGRIDS' most advanced format. It keeps colour and point density high, so the splat stays accurate and photorealistic. The LCC pipeline also connects cleanly to SketchUp. With Supersplat and other routes, the XYZ coordinates often need correcting.
Before import, we clean the scan: excess furniture, people who walked through, moving vehicles, and any part of the site the project doesn't need.
File sizes are more manageable than people expect. A 10-minute scan usually lands between 100 and 300 MB, and even 40-minute scans tend to stay under 1 GB. Files that size don't slow SketchUp down.
The Gaussian Splatting Manager walks you through setup in six steps:
Cropping
Orientating
Aligning
Scaling
Positioning
Saving
Our first move is always to crop to about 90%, so we can focus on the core of the scan while we align it. Once it's positioned, scaled and squared to the axes, we take the crop back to 100%. The surroundings and distant reflections return, and they make a beautiful backdrop for presentation slides later.
The most common problem isn't the plugin. It's the file. A client sends a PLY that turns out to be a plain point cloud, not a Gaussian splat. The test is quick: open it with Scan Essentials instead. If it loads there but not in the splat plugin, you know what to ask the client for.
Scale, axes and trust
Axes. We align the scan's X-axis to something straight and reliable: a wall, a floor tile line, a pavement edge. Then we place the design and the scan together on that axis.
We also have a philosophy about the origin: X, Y, Z 0 sits at the centre of the project. Everything in SketchUp works outward from the middle, and your scenes and LayOut sheets line up far more neatly for it.
Scale. We use a triangle of control points. On arrival, we look for three distinct existing features we can measure between in height, width and depth. If the space doesn't offer three, we place scanning targets as control points.
Then we measure A to B, B to C and C to A twice: once with a Stanley tape on site, once in the scan data in LCC Studio. If the numbers disagree, we know which direction the scan needs correcting, X, Y or Z.
Accuracy. A splat is accurate to roughly 12–20 mm. That's plenty for schematic design. It isn't enough for everything. We re-measure by hand:
windows and door frames
artworks
floor-to-ceiling heights and crawl spaces
HVAC, fire alarms and sprinkler systems
anything we're designing that's under a metre
The tolerance trap
This part catches people.
Say the splat tells you a cavity is 5 m. You design the fixture at 4.96 m, leaving 20 mm either side. Sensible.
But what if the scan's own error has already eaten that margin? What if the cavity is actually 4.96 m? You've just designed a fixture exactly the size of the hole it's meant to fit into.
Your tolerance has to be an interpolation, not an extrapolation. It has to sit inside the measured reality, not stack on top of the scan's uncertainty.
We felt this on the Jameson Distillery on Tour project at the Castle of Good Hope. The archways you walk through are historic and irregular, and we were building a steel tunnel structure through them. The scan's error margin and our design tolerance compounded, a double negative, in exactly the place where steel doesn't forgive.
That's why we agree up front with the shopfitter that they site-measure every fixture after design approval, before manufacturing. The splat gets us through concept and schematic design. The tape gets the fixture onto site.
Why we model inside the splat
We model directly against the splat, in it and around it, never in a separate file.
The reason is contextual modelling. When a client says, "That fixture goes in the top-left corner over there," a photo set sends you scrolling for the right frame. A 360 video sends you scrubbing for the right moment. In a splat, you simply fly there.
(When splats go 4D, we'll be able to play back a site with its people moving through it, and study how they flow. That's for another post.)
Cutting what's being removed. We cut holes in the splat rather than living with ghosts. The scissor icon in the Gaussian Splatting toolbar, Cut Selection as a new entity, lifts the selection into its own entity in the manager, where you can hide or unhide it.
Nothing is deleted. That matters, because you can always check yourself: I removed this, but is it structural? Does it need to come back before I sign off with the client?
Baked-in lighting. A splat carries the light it was scanned in. Your model doesn't. At concept stage that doesn't matter. Down the line, I expect tools that strip the lighting out, so you can relight splat and model together in V-Ray or D5 and test scenarios properly. We're not there yet.
Working views. Day to day:
Reduced opacity when aligning.
Point cloud mode when aligning, or when we're drawing out existing elements.
Full splat at reduced opacity when balancing the whole composition.
We rarely use the drawing style. It renders individual points as aggressive oval blotches, and even with lower opacity it's just not a helpful way to model.
Tags and scenes. We organise every file to work with or without the splat. Some scenes show it and some don't, so a technical drawing isn't cluttered with surroundings it doesn't need.
Deliverables
Gaussian splatting is still new to most clients, and the richness of the data can overwhelm. So we present it ourselves, in person or on a call, live in SketchUp. We can orbit the model, switch tags and scenes, toggle the splat and change its opacity. Nothing else offers that flexibility.
Renders, walkthroughs and LayOut drawings come later, once the client has seen the environment once or twice. For a new client, an early PDF that feels deliberately sketchy, light on text, just materials and finishes, is often enough to agree a layout.
Is a rendered view better than the splat with SketchUp elements over it? V-Ray 7, D5 and AI enhancement all have their place. But none of them reliably reads to the client as the better way to see the design. The splat with the model over it is usually enough.
Landlords are the exception. Mall landlords and brand teams generally won't approve off a scan. They want the as-built drawn out. So we draw it, tracing off the splat, and compare it against whatever old plans the mall manager has shared. Those rarely match.
Brand teams are different again. They seldom read line weights and line types. A scan speaks their language.
Where the splat misleads you
Drift, in long scans of repetitive spaces. Control points pull it back.
Glass and mirrors. Matterport handles this well: before processing, you mark on the map where the windows and mirrors are, and the software accounts for them. A Gaussian splat captures hard surfaces, not transparent ones. Look through a window and you get a hazy, ghostly patch where the glass should be. It makes the environment read a little disruptively.
Geo-referencing. A splat is a reality-capture scan, not a survey. If a project needs to be geo-referenced, the surveyor's set-out points are the single source of truth. Some scanners with an RTK module and a base station can position a scan properly, but not all of them, and only outdoors.
Beyond that? I honestly struggle to name something I'd never trust a splat for. It replaces the side plan, the old CAD plan, the reference photos and the Match Photo sessions. Most of that becomes obsolete.
The time
On site, four hours of hand measuring becomes about one hour of scanning.
Back in the studio, one hour of placing and checking the splat replaces four hours of redrawing the existing site, and up to twenty hours when detailed drawings are needed.
And that gap is about to widen. Within the next six to nine months, I expect AI tools to redraw existing conditions straight from a point cloud, a splat or a 360 video.
If Trimble asked
I'd merge Scan Essentials and the Gaussian Splatting plugin into one extension, build in 360 video tools, and connect Trimble ProjectSight, which already supports 360 capture, so walkthroughs can be positioned inside the SketchUp model and open in a 360 viewer on click.
Where to start
Start with an object. Put a lamp, a bottle, a stack of books or a cardboard box on the boardroom table and scan it in Scaniverse. Then scan the whole boardroom. Then the walkway outside it. Then your office entrance.
Move from your phone to an iPad Pro for longer scans. If it still interests you, buy an Insta360 X6 and use their processing software for a while. (I can't yet confirm it exports a PLY, but I expect it will.) After three to six months with clients, step up to a dedicated scanner, whatever you can get your hands on.
For designing inside a splat, that usually means a handheld SLAM scanner: XGRIDS, NavVis, a Leica BLK2GO. They do differ, but it's like choosing between an iPhone 17 and a Galaxy S26. The real differences are the software and what they charge you to process.
A tripod scanner like the FARO Focus or the Trimble X7 and X9 is a different tool for a different job. It's slower to move around a site, but more precise, and it's what you reach for when the measurement matters more than the walkthrough.
That's your starting point.
The splat isn't the drawing. It's the room you draw in.
#GaussianSplatting #RealityCapture #SketchUp #RetailDesign #DesignWorkflow
