Phone to photoreal 3D

Scanverse — phone to photoreal 3D building scans

Walk around a building. Get the building.

StatusIn build
ClientOwn R&D
Shipped2026
TypeTool

What it is

Record a building exterior on a phone, and Gaussian splatting rebuilds it as a photoreal fly-through. Capture and viewing happen on the phone; the heavy reconstruction runs on a GPU — the same split every commercial scanner uses, stated plainly instead of implied away.

The numbers

Splats in the test build: ~20,000
Runs on device: Capture + view
Runs on GPU: Reconstruction
Milestone: Viewer shipped

What is inside

Guided orbit capture so the reconstruction actually converges
Mobile-first viewer with auto-orbit and reset
Pairs with the 3D twin in PROP AI

In more detail

What happens between walking the building and seeing it

Five stages. Orbit capture and frame extraction happen on the phone, and blurred frames are dropped before anything is uploaded. The camera solve and the splat training happen on a GPU server. The trained scene then comes back to the phone, where it renders in the browser at sixty frames a second.

Why the walk around the building has to be guided

An uneven orbit is the usual reason a reconstruction fails, so the capture guide is designed to hold you at a steady radius for a full circle rather than let you drift closer and further. Frame selection then drops anything with motion blur on the device, before anything is uploaded, so bad images never reach the camera solve.

Why the reconstruction cannot run on the phone

Structure from motion has to work out where every photograph was taken from, and the training that follows is millions of gradient steps fitting oriented gaussians to those images. On-device training of a whole building is not feasible today. Every commercial scanner splits the work the same way; here the split is written on the page rather than implied away.

Where a finished scan is meant to be used

It pairs with the three-dimensional building twin in JD PROP AI, where a scanned facade can stand in place of a modelled block. What comes off the GPU is an exported splat file — .ply or .ksplat — and that is what the viewer loads. The viewer is built mobile-first, with auto-orbit and a reset control, and the test build carries about twenty thousand splats.

Built with

Vite Three.js gaussian-splats-3d COLMAP gsplat

Open the live demo Talk to us on WhatsApp

It boots in your browser and runs the real logic — no form, no sales call.

Common questions

Is Scanverse finished and available?

Not yet — it is in build. The viewer is the milestone that shipped: it loads a trained scene and orbits it on a phone. Guided capture and the reconstruction queue are the pieces still being built, and it is our own research rather than client work.

What do I need to capture a building?

A phone and a clear path all the way around the building. That is the capture stage as designed: the guide holds a steady radius for one full circle, and only sharp frames are uploaded. Guided capture is one of the pieces still being built — the viewer is the part that shipped. Everything heavy happens afterwards on a GPU, not while you are standing outside.

How is a scan different from the 3D building already in JD PROP AI?

The one in PROP AI is a modelled block — geometry drawn to stand for a building. A Scanverse capture is a photoreal reconstruction of the building that is actually there, colour and all, and it is designed to drop into that same twin in place of the model.

Does a scan open in a browser, or does it need an app?

It opens in a browser. The viewer is a web page built mobile-first, so a scan can be shared as a link and orbited on the phone that receives it. Nothing has to be installed at either end.

More of the work

JD Software Solutions — a four-person software studio in the United Arab Emirates. We build the system, hand over the code, and stay on to run it.

All of the work