SplatEngine Start a capture

How it works

From a camera on site to a line in your page

Gaussian splatting reconstructs a scene as millions of soft, coloured ellipses rather than a mesh with textures on it. Each one stores position, size, orientation, opacity and how its colour changes with viewing angle — which is why glass, chrome and foliage survive. Here is the actual pipeline, and then the honest limits.

The pipeline

Five stages. You’re involved in two of them.

Typical turnaround is five to ten working days from capture to live embed, depending on scene size and how many review rounds you want.

01

Capture on site

We shoot overlapping stills or video, covering every surface from multiple heights and angles. Coverage matters more than camera quality: a corner nobody photographed becomes a smeared, cloudy artefact later. Interiors take two to four hours, objects around forty-five minutes on a turntable rig. We need the space still and the lighting constant, so no moving people in shot and no curtains opening halfway through.

02

Solve the camera positions

Every frame is matched against every other to work out where the camera was when each shot was taken. This produces a sparse point cloud and a set of camera poses. Blurry frames, duplicates and anything that moved get dropped here.

03

Train the splats

The point cloud is optimised into millions of Gaussians. The model renders its guess, compares it to your real photographs, and corrects itself — tens of thousands of times. Splats split where detail is needed and are deleted where they contribute nothing. This is the GPU-heavy part and runs for a few hours.

04

Clean and optimise

We cut floaters and background junk, crop to the area you care about, add waypoints and hotspots, then compress the scene and build level-of-detail steps so it streams sensibly on a phone. You review a private link here and tell us what to fix.

05

Publish the embed

The approved scene goes to our CDN with a permanent URL. You paste one iframe. If you later want a re-capture or an edit, we swap the scene behind the same URL and your site needs no changes.

Hardware & limitations

What it needs to run, and what it won’t do.

We would rather lose the job than ship something that stutters on your customers’ phones. Read this part properly.

Device classExperienceScene budget we target
Desktop, dedicated GPUFull quality, 60fps+Up to 4M splats
Laptop, integrated graphicsFull quality, 45–60fpsUp to 2.5M splats
Phone or tablet, last 4 yearsReduced detail, 30–60fps1–1.5M splats streamed
Phone, 5+ years oldLow detail, may stutterUnder 800K splats
No WebGL2 availableFalls back to stills

Limit 01

It is not a measurement

A splat looks correct but is not a survey. Do not scale drawings, order materials or price a job from one. If you need dimensions, you need a laser scan or a proper survey — we can work alongside one.

Limit 02

Everything must hold still

People walking through, traffic, running water, moving foliage and changing daylight all appear as smears or ghosts. We capture out of hours, or we crop the problem out.

Limit 03

Uncaptured angles look wrong

Splats only know what a camera saw. Under a low sofa, behind a wardrobe or above head height on a tall ceiling will look soft or cloudy unless we could physically get the camera there.

Limit 04

File size is real

Scenes run from roughly 20 MB to over 100 MB. We stream them progressively so something appears within a second or two, but a full-detail scene on poor mobile data still takes time.

Limit 05

Lighting is baked in

The scene keeps the light it was captured in. You cannot switch the lamps on afterwards or move to golden hour. If lighting matters, we plan the capture around it.

Limit 06

Locked-down browsers

Some corporate and public-sector environments disable WebGL2 or block third-party iframes. We provide a still-image fallback and can supply a direct link for those users.

Battery, heat and long sessions

Real-time 3D works the GPU, so phones warm up and drain faster than on a normal page. We cap the frame rate, pause rendering when the frame scrolls out of view, and drop detail when a device reports thermal pressure. It is still worth placing the embed where people choose to engage with it rather than autoplaying it at the top of every page.

Your side

What we need from you

  • Access to the space with nobody moving through it
  • Lights on and set the way you want them remembered
  • The space dressed and tidied — we capture what is there
  • One person who can approve the draft
  • Somewhere on your site the embed will actually live

Our side

What we handle

  • Capture, training, cleanup and optimisation
  • Hosting, CDN delivery and uptime
  • Device detection and quality fallbacks
  • Viewer branding, waypoints and hotspots
  • Scene updates behind the same embed URL

Next step

Send us a space, get back an embed link.

Tell us what you want captured and how people will use it. We’ll come back with a scope, a price and a realistic timeline — usually within two working days.