Gaussian splatting for real estate: 360 camera, LiDAR or Matterport, and where to compute the model

Quick answer

Gaussian splatting (3DGS) turns photos or video of a home into a 3D scene that viewers fly through smoothly in a browser, instead of jumping between panorama points. For interiors, a 360 camera and a video walkthrough give the best quality for the cost. A LiDAR phone is the fastest but weaker on larger spaces. A Matterport-style scanner gives measurements and a proven tour, and a drone shows the building from outside. You can compute the model with open tools on your own GPU or in the cloud. A splat is not a measuring tool: glass, mirrors and motion break the model, and the geometric error reaches centimetres. For floor area you need LiDAR or a measurement.

What 3D Gaussian splatting is, in plain words

Picture a scene made of one to several million small, semi-transparent, coloured "droplets" shaped like ellipsoids. Each has a position, size, rotation, opacity and a colour that can change with the viewing angle. That is why the reflection on the floor moves when you move the camera. The algorithm keeps adjusting these droplets until the image seen from each photo's position looks like that photo.

Kerbl et al. described the method in 2023. According to the paper, it renders in real time (at least 30 frames per second at 1080p) with quality on par with the best neural methods. The starting point is always a set of photos with known camera positions, computed with Structure-from-Motion, most often in COLMAP.

The two methods a splat is usually compared with:

• A photogrammetry mesh. A classic 3D model of triangles and textures. Good for measurement and CAD, but weak on reflections, glass, vegetation and thin details. It looks like a model, not like a photo.
• NeRF (Mildenhall et al., 2020). A scene stored in a neural network that computes the image pixel by pixel. Photorealistic, but rendering is slow, so it is hardly used in browsers.

What to capture with: four options

A 360 camera (e.g. Insta360-class)

One recording captures the whole surroundings, so by walking slowly through the home you collect frames from many positions without worrying about framing. The Insta360 X6, launched in August 2026, records 8K 360 video at 50 fps on two 1/1.1-inch sensors. You do not need the newest model for splats, though. A slow pace, stable exposure and covering each room from a few heights matter more.

There are two processing routes. The first is ready-made services: the Insta360 app has a Spatial Capture feature (for the X6, X5 and X4 in selected modes), and since April 2026 Insta360 has partnered with Splatica, a platform that turns 360 footage into 3DGS scenes. Both process the footage in an outside provider's cloud. According to the Spatial Capture terms reported by Radiance Fields, reconstruction is done by a company based in Shenzhen. The second route is your own pipeline: since version 3.12 COLMAP has a Structure-from-Motion example for 360 panoramas that splits spherical frames into overlapping perspective views, and since version 4.1 it also supports spherical (equirectangular) cameras natively.

A phone with LiDAR

A Pro iPhone or an iPad Pro with Scaniverse or Polycam is the quickest start. Scaniverse (Niantic Spatial) computes splats on the device without sending data anywhere, and Polycam exports splats to PLY. Phone LiDAR gives you a splat at metric scale. A 2026 ISPRS study showed, however, that phone LiDAR reconstructions drift, and the drift grows with the size of the object: when the building perimeter doubled, the error grew 1.5-2 times. Reconstructions based on photos alone did better both visually and spatially.

A Matterport-style scanner

The Matterport Pro3 combines LiDAR with panoramas. The manufacturer states an accuracy of ±20 mm at 10 m, a range of up to 100 m (in E57 export) and under 20 seconds per scan point. It is the market standard for virtual tours and a source of floor plans and measurements. The Matterport tour itself, however, is panoramas at scan points plus a 3D mesh, not a splat, so movement jumps from point to point. Images from such a scan can feed a splat, but that is a separate process.

A drone for the building and its surroundings

To show the building, the plot and the neighbourhood, a drone works best. DJI Terra has computed 3DGS models from drone images since version 5. Since July 2025, Zillow has shown SkyTour on selected listings: an exterior view of a home built with Gaussian splatting from drone footage. Flights must follow aviation rules and permits, which can be a constraint in built-up areas.

The times in the table are rough estimates for a single apartment, not measured results. They depend on the number of rooms and how many shots need to be repeated.

Where and how to compute the model

The pipeline looks similar whatever the tool:

1. Frame selection. Extract sharp frames every few centimetres of movement from the video and drop blurred ones.
2. Camera poses (SfM). COLMAP computes the position of every frame and a sparse point cloud. Version 4.0 (March 2026) integrated the global mapper from the GLOMAP project, which its authors report is 1-2 orders of magnitude faster than the classic approach.
3. Splat training. Usually tens of thousands of iterations.
4. Cleanup. Remove floaters (loose Gaussians hanging in the air), crop the view beyond the windows and level the floor, for example in the SuperSplat editor.
5. Compression and publishing. Convert to a web format and embed it in a page.

Tools that are available and maintained as of September 2026:

Nerfstudio with its splatfacto method still works, but the main repository has had no changes since July 2025, and development has moved to the gsplat library. Inria's implementation is licensed for research only, so for commercial jobs pick tools under Apache, BSD or MIT, or check the terms of GPL and AGPL.

GPU memory. Inria's implementation needs 24 GB of VRAM to reach the quality reported in the paper. gsplat claims up to 4 times less memory with shorter training time. Smaller scenes, such as a single apartment, can often be computed on a consumer card, while large houses or whole office floors need more memory, fewer Gaussians or a split scene.

Cloud or local. Cloud services are convenient: you upload the footage and get a link a few hours later. Your own pipeline needs a GPU and some work, but gives you control over quality, format and data. That matters, because footage from inside a home is often personal data: photos on the walls, documents on a desk, faces of the residents. Before sending footage to a service, check where it is processed and how long it is kept. A rented cloud GPU in an EU region is a middle ground.

Viewing on the web

Splat viewers. The most mature open options are the PlayCanvas engine with the SuperSplat editor (both MIT) and Spark from World Labs, a Three.js renderer built on WebGL2 that its authors say runs on almost any device, including low-powered phones. Brush also has a web viewer, but it is built on WebGPU, so for now it works only in Chrome and Edge. The simple antimatter15 viewer, which popularised the .splat format, now points users to Spark itself.

File size. In the original PLY each Gaussian is 59 floating-point numbers (position, scale, rotation, opacity, colour and spherical harmonic coefficients), or about 236 bytes. One million Gaussians is about 236 MB, and PlayCanvas notes that PLY scenes range from 50 MB to several GB.

SOG stores Gaussian attributes as WebP images plus a meta.json file. A similar idea, arranging Gaussians into 2D grids and compressing them like images, was described by Fraunhofer HHI researchers in Self-Organizing Gaussians (ECCV 2024): in their example a 174 MB scene fit into 17 MB at the same quality. The open splat-transform tool converts between formats.

Phones. Splats are expensive mainly because of per-pixel work (fill rate), and every Gaussian is sorted by depth every frame. PlayCanvas recommends a budget of about one million Gaussians for mobile devices and three million or more for desktop, turning off anti-aliasing, and lowering render resolution on weaker devices. For large scenes there is Streamed SOG, which loads detail based on camera distance.

A practical workflow for a listing

1. Preparation. Tidy up, switch on every light, and treat the blinds consistently (all open or all closed). Remove documents and family photos from view. Keep pets in another room.
2. Capture. A 360 camera on a pole at about 1.5 m, a slow walk, going around each room along the walls and then through the middle. Do a second pass low down to catch furniture from below. Record the transitions between rooms without stopping, because they connect the scene.
3. Dimensions separately. If the listing states floor area or includes a floor plan, take it with a laser measure, a LiDAR scanner or from the documentation. The splat is the visual layer.
4. Compute. Frames into COLMAP (panorama mode), training in gsplat, Brush or LichtFeld Studio, and a quality check from several viewpoints.
5. Cleanup. In SuperSplat, remove floaters, crop the view beyond the windows and set the camera's starting point.
6. Publish. Convert to SOG or SPZ, test on a cheap phone, and embed it in the listing page with PlayCanvas or Spark. Keep ordinary photos and a floor plan next to it, because not everyone wants to fly a camera around.

Limits to explain to the client

Glass and mirrors. Standard 3DGS does not model reflections physically: a mirror usually becomes a "window" into a second, fake room behind the wall, and large glazing gives blurry or patchy surfaces. Research work such as Mirror-3DGS and MirrorGaussian (2024) adds explicit modelling of the mirror plane, but that is not yet standard in the tools. LiDAR struggles with these surfaces too and returns missing or wrong readings.

Motion. The method assumes a static scene. People, pets, curtains in a draught or changing light cause smears and floaters. Record an empty property under constant lighting.

Accuracy. A splat is optimised for appearance, and geometry is a by-product. In the 2026 ISPRS study, photo-based reconstructions had a target error of 1-2 cm, splatting itself changed it by less than 8 mm, but it added floaters, and the authors concluded that building splatting is not yet a mapping tool. In interior tests described by Lidar News, the mean geometric error of splats was 7-8 cm. Without reference points, the model also does not know its own scale.

Checklist

1. Is the goal marketing and a sense of space rather than measurement? Then a splat makes sense.
2. Do you have a 360 camera or a LiDAR phone and an hour for the capture?
3. Do you know where the model will be computed and whether the footage may leave the EU?
4. Do the dimensions and floor plan come from a source other than the splat?
5. Were mirrors, large glazing and motion kept to a minimum during capture?
6. Does the compressed file load in a few seconds on a cheap phone?
7. Do the tool licences allow commercial use?

Sources

• Kerbl et al.: 3D Gaussian Splatting for Real-Time Radiance Field Rendering (2023) and Inria's repository with its licence
• Mildenhall et al.: NeRF, Representing Scenes as Neural Radiance Fields for View Synthesis (2020)
• McNally et al.: A Comprehensive Evaluation of the Spatial Accuracy of Building Gaussian Splatting, ISPRS Annals (2026)
• Lidar News: Gaussian Splatting and LiDAR, A Practitioner's Field Guide (2026)
• Morgenstern et al.: Compact 3D Scene Representation via Self-Organizing Gaussian Grids (ECCV 2024)
• Meng et al.: Mirror-3DGS (2024) and Liu et al.: MirrorGaussian (ECCV 2024)
• COLMAP and its panorama SfM example
• GLOMAP, migrated into COLMAP
• gsplat, Brush, LichtFeld Studio, OpenSplat, nerfstudio
• Jawset Postshot
• PlayCanvas: Recommended Tools
• PlayCanvas: SOG format and PLY format
• PlayCanvas: Performance
• PlayCanvas engine, SuperSplat, splat-transform
• Spark, antimatter15/splat
• Niantic SPZ
• Khronos: KHR_gaussian_splatting
• Matterport Pro3: specifications
• DJI Terra V5.1
• Lidar News: Zillow SkyTour Uses Gaussian Splats
• Newsshooter: Insta360 X6
• Radiance Fields: Insta360 Spatial Capture on the X6
• Insta360: Splatica partnership
• Scaniverse on the App Store
• Polycam: exporting splats