Gaussian splatting for real estate: what a 3D apartment finder can honestly show

Property marketing already uses 3D: interior scans, CGI flythroughs and virtual tours. Gaussian splatting adds something those often lack: the real site, photo-real, built from ordinary drone footage, viewable in any browser, and cheap enough to remake whenever the site changes. Unlike a textured mesh, a splat keeps trees, glass and sky looking natural, which is most of what people look at when they judge a location.

To see what that is worth, we built a full apartment finder on top of one. It is live as Simam Spatial Residences, with the method on the lab page.

What the 3D itself can tell a buyer

• Heights and floors. Each building is measured in the model: the main tower is 34 m with 10 floors, ±12% because scale comes from counted storeys.
• Outlook. A unit's view is shown as a wedge across the real 3D: park side or city side, and what sits in between.
• Surroundings, with real distances. The park, the plaza, the rail line, the car park and the elevated road are all in the footage, so their walking times are measured, not guessed.
• Things worth checking. AI search over the model found a freight line about 30 m west of the tower. That is exactly the kind of fact a brochure leaves out and a buyer wants to know.

What has to be supplied

The 3D cannot know prices, availability or what is inside. In the demo these are sample data, marked "sample" throughout.

• A stacking plan: every unit coloured on the facade by availability, with filters for bedrooms, outlook, budget and available only. In a real deployment this comes from the sales system.
• Listings: price, size, level, features, a floor plan and a viewing request.
• Interiors: we used Gemini to make eight interior and amenity pictures for $0.55. Each picture was drawn on a real crop of the drone footage, so the windows show the actual park or skyline. A developer would use photography or renders from the real plans.
• Places nobody filmed: the playground, shops, grocery, station and school are mock-ups, labelled as samples.

What we tried and dropped

A true "view from the window" looked into blur, because the drone never filmed outwards from the tower. A straight-down bird's-eye view showed sky-coloured smears where the footage never reached. Both were dropped in favour of viewpoints the drone actually covered: aerial, from the park looking up, three-quarter, and a slow orbit. The lesson for anyone selling with splats is to keep the camera where the capture was, and to plan the capture around the views buyers will want.

Raw captures also tend to look grey and flat. A light colour grade on the model (a little more exposure, contrast and saturation) made the site look as it does in person, without touching the overlays.

Labelling is part of the product

Property advertising has to be accurate, and computer-generated images must not mislead. In our viewer every AI picture carries an "AI picture · sample" label, sample data is marked, and measured figures carry their error. Labels like these don't weaken the experience. They are why a buyer can trust the parts that are measured.

What a developer would need

• Their own drone flight around the site, all sides, ideally on an overcast day.
• A live availability feed from the sales system.
• Real interiors: photography, or renders from the floor plans.
• Surveyed points if exact heights or distances are going into contracts.

The model itself took under ten minutes to make. The steps and timings are in from drone video to a 3D model in under 10 minutes.