Gaussian Splatting vs Photogrammetry vs LiDAR: What a Designer Actually Needs
Part 10 of our Gaussian Splatting Series expanding the comparison from Part 1 into a standalone guide, with a cost ladder from phone to K2.
1. The three technologies, properly defined
When we look at what a designer actually needs, we're dealing with three types of technology that each serve a different purpose:
LiDAR is, at base level, a laser mounted in a handheld device, on a tripod, or on a vehicle-mounted mobile station that scans proportionate geometry in a static condition.
Photogrammetry is the use of a digital camera to create a 3D model: the computer understands and measures depth within 2D photographs, then augments them into a 3D model built from the photographic textures and the depth information inside the photos.
Gaussian splatting is essentially the combination of the two: the LiDAR scan result laser points, which can be colorised with the photogrammetry data laid on top. A highly accurate, photographic environment.
As appealing as the splat is, there are situations where you would choose one of the others. Let's take each in turn.
2. LiDAR: when the millimetre matters
Depending on the scanner you choose, LiDAR can provide an accuracy that is almost a true reflection of real-life conditions one to two millimetres with a total station. That's the accuracy required in industry and by government organisations, and certain projects demand even more. Within that context, LiDAR provides the most accurate result, full stop.
Point cloud density can be so high that a colorised point cloud almost reads as a photograph. But it is still points and here's a useful way to think about density: treat it like the DPI of printed material.
A high-quality total station like a Leica RTC360 scans at about a million points per square metre, with an average point spacing of a millimetre. If you used those points as a render, the equivalent print resolution would be roughly 25 DPI. Compare that to the world of print: home printers output at 300–600 DPI, scanners run 300 to 1200. As a rendering, a standard LiDAR scan is coarse.
Go to ultra-high terrestrial laser scanning 10 million points per square metre at 0.3 mm spacing and you reach about 80 DPI. Since 72 DPI is the floor people accept for shared low-quality PDFs, there's a case for an ultra-high point cloud carrying an A4 or A3 printed floor plan.
And that, in a sense, is where LiDAR ends: unbeatable for measurement, limited as an image.
3. Photogrammetry: when the photograph matters
Now take a 50-megapixel DSLR and scan a building in a chosen lighting condition say, golden hour with a photo set of around 10,000 images. In a tool like Postshot, those photos are arranged into a 3D model using the AI's shadow and depth sensing, based on each photo and the direction it was taken from.
There's no LiDAR data here, so there's no one-millimetre accuracy only the depth read from field of view, depth of field, and the casting of shadows within the photographed object. But the photorealistic result, given a dense photo set and a good camera, is phenomenal.
The catch is labour and light. Taking 10,000 photos thoroughly takes three to four hours, and the lighting conditions change while you work. Without adjustment in post-processing, one side of the building looks like early morning and the other side looks closer to lunchtime.
Where photogrammetry truly shines is the static, aesthetic interior a museum. The resolution density gets high enough to build a VR environment where you can actually read the placards alongside the objects on display. That is a very effective use case.
4. Gaussian splatting: when you need both
Combine the LiDAR data with the photogrammetry through one hardware device and you get an extremely detailed, photographically dense model: accurate enough to measure from for construction at 12–20 mm, with photographic quality that lets you read real-life conditions down to 2–5 centimetres of detail the spine of a book in a library, on a high-detail scan.
5. So when do you use which?
Photogrammetry: static interiors, where photographic fidelity is everything.
Gaussian splatting: static or dynamic interiors and exteriors, where you need to design, present, and measure in one environment.
LiDAR: ultra-high-precision scans for surveying and quantity surveying.
6. The cost ladder: from phone to K2
~$500 a phone. Where everyone should start (see Part 9, the phone tutorial).
Under $1,000 phone + Insta360 X5. The X5 retails around $550–600; add a tripod, lights, the invisible selfie stick or a gimbal to offset movement, and you've packaged a wonderful step-in solution for speed capture.
~$1,000+ iPad Pro. A slightly higher price point, but now you have photogrammetry and LiDAR in one device, and with an M3 or M4 chip, more processing control over your data than a phone gives you.
$2,500–4,000 Matterport, Giraffe360. Scanning equipment built on photogrammetry and 360 photography with multi-pass HDR. Incredible photorealistic results not necessarily the best Gaussian splats, but phenomenal photogrammetry and medium-quality LiDAR.
$6,000–10,000 Lixel K1/K2, PortalCam, and competitors. This is where the hardware does both really well: proper LiDAR and proper photographic capture in one handheld SLAM device.
The ladder matches your business. In the beginning, you're testing demoing the technology as part of your site surveying, starting with the phone or iPad, maybe augmented with the 360 camera for speed, still under a thousand dollars. As clients increasingly desire this technology, make it part of their requirements, and your workflow grows, the quality of your hardware can climb alongside the price point of your services.
7. The hybrid argument
At what point do you use one technique over another? Often the real scale is quality versus time not price. Price becomes negligible once you're doing 50 to 100 projects a year and can charge a minimum fee for the service.
Run that maths: if you need 50 to 100 scans a year at a consistent retail/shopfitting level of quality, picking up a K2 and scanning a store in 20 to 30 minutes is a far more considered decision than taking a Matterport Pro2 or Pro3 and standing in that premises for one to two hours. The hardware's upfront cost does not equal the continuous labour fee of people scanning with a slower device.
And then there's the hybrid workflow we actually run. When we navigate a space with a client, we'll often have an Insta360 in hand, recording a 360 video while both of us talk through the project. It's a note-taking tool, a site surveying tool, and a meeting-capture tool at the same time. Afterwards, we can extract that video and process it into a low-level Gaussian splat through a Lixel tool or a third party.
From there, the hybrid ladder kicks in: after the meeting, we bring out the hardware renowned for what the project needs the DSLR for photogrammetry, the total station for LiDAR accuracy, or the K2 for the combination.
And sometimes the hybrid collapses into one visit: we take the K2 to the client meeting and scan while we walk, using a tool like Plaud AI, Nothing's Essential Space, or Granola to record the conversation and generate the meeting notes. One walk through the site: the meeting captured, the notes written, the site scanned.
That ends part 10 of our series on Gaussian Splatting. I hope this gives you insight into how to start using splatting hardware in your business, at whatever rung of the cost ladder you're comfortable with as you try the technology out. For now — a wonderful week ahead. Take a look at our YouTube videos and previous posts on this topic, have a look at our current projects around the website, and feel free to leave comments or questions below.
Next post: What is SuperSplat, and how do you clean up a splat?
What is Gaussian Splatting, Part 9 Phone Scan to SketchUp
