Definition: Gaussian splatting (more precisely: 3D Gaussian Splatting, 3DGS) is a method that reconstructs a photorealistic 3D scene from ordinary photos or videos. Instead of a triangle mesh, the scene consists of millions of small, semi-transparent ellipsoids ("Gaussians"). Together, they reproduce the appearance of a real object or room. Researchers around Bernhard Kerbl (Inria, Université Côte d’Azur and Max Planck Institute for Informatics) presented the method at SIGGRAPH in 2023. It achieves display in real time.

In B2B mechanical engineering, Gaussian splatting is used for: documentation of production halls and existing systems, virtual tours through plants, archiving of trade show booths, and photorealistic backdrops for presentations on the web, in AR and in VR.

Advantage: Capture real environments in photorealistic quality, with a camera instead of complex 3D modeling.

English: Gaussian Splatting, 3D Gaussian Splatting (3DGS)
German: Gaussian Splatting, Gauß-Splatting

How It Works

From photos to a scene

  1. Capture: You photograph or film an object or room from many angles, with a smartphone, a system camera or a drone.
  2. Determine the camera positions: A structure-from-motion method (for example, COLMAP) calculates where each image was taken and makes a sparse point cloud.
  3. Initialization: Each point becomes the starting point of a Gaussian.
  4. Optimization: On the GPU, the software adjusts the position, orientation, size, opacity and color of all Gaussians, so that the rendered views match the photos as closely as possible. Where details are missing, it splits or adds Gaussians and removes unnecessary ones.
  5. Result: A scene of often several million Gaussians.

Tips for the capture: Even, diffuse light, a high overlap between the images and a steady, continuous camera movement improve the result considerably. Avoid moving people, running conveyor belts and changing screen content during the capture, because they appear as streaks in the scene.

What a Gaussian stores

  • Position: Center point in space
  • Rotation and scale: Shape and orientation of the ellipsoid
  • Opacity: How much it hides what is behind it
  • Color: As spherical harmonics, so that the gloss changes with the viewing angle

Display in real time

For rendering, the Gaussians are sorted by depth and projected onto the image ("splatted"). This rasterization is much faster than sampling neural networks. The original paper achieved at least 30 frames per second at 1080p. Today, splats also run in the browser and on headsets, see real-time rendering.

Comparison: Photogrammetry, NeRF and Gaussian Splatting

Criterion Photogrammetry (mesh) NeRF Gaussian Splatting
Data structure Triangle mesh with texture Neural network Point-based Gaussians
Visual quality Good, weak with fine structures Very photorealistic Very photorealistic
Rendering speed Real time Usually slow Real time
Editing In any 3D program Hardly possible Cropping, cleanup, limited
Clean surfaces Yes No No

Photogrammetry supplies meshes that you can use in conventional tools, but it often looks messy for cables, grids or reflective surfaces. NeRF (Neural Radiance Fields, 2020) makes impressive images, but needs a lot of computing power. Gaussian splatting combines the image quality of neural methods with the speed of conventional rasterization.

Difference from LiDAR: LiDAR measures distances directly and supplies dimensionally accurate point clouds. Gaussian splatting optimizes for appearance. For as-built surveys with requirements for accurate dimensions, the two methods complement each other.

Status in 2026: Standards and Tools

For a long time, each tool stored splats in its own formats, usually as large PLY files. Now there are open standards:

  • glTF: In February 2026, the Khronos Group published a release candidate of the extension KHR_gaussian_splatting. It is now ratified and listed in the official Extension Registry. It stores splats as point primitives. Older viewers can show them as a point cloud instead.
  • OpenUSD: Version 26.03 adds its own schema for Gaussian splats (ParticleField).
  • Compression: Formats such as SOG and SPZ make the file size many times smaller than PLY. This is necessary for mobile devices and the web.
  • Tools: Browser-based editors such as SuperSplat (open source) clean up, crop and export splats.

B2B Applications

Sales: show reference systems

A manufacturer captures a running filling system at a customer site and shows it to prospects in the browser or on the tablet, photorealistic and with the surroundings. This does not replace a plant visit, but it makes the size and the installation situation tangible.

Trade show: archive and reuse booths

You capture a trade show booth on the last day. The result serves as a virtual tour for customers who could not travel to the show, and as a reference for planning the next trade show presence. You can also capture exhibits that have no CAD model in this way, for example older machines or samples.

Service: as-built documentation

Before a conversion, you record the situation in the hall: pipes and cables, access routes, neighboring machines. Service technicians see in advance what they will find on site, and plan their tools and access routes.

Training: realistic environment

VR training becomes more realistic when the virtual workplace matches the real hall. The splat scene is the background. The machine that you operate stays a conventional 3D model.

Design review: new system in an existing hall

You place a virtual prototype in the captured hall. This lets planners check how the new line looks in the real environment: in VR, on a screen or with augmented reality.

Limitations

  • No replacement for CAD: Splats contain no surfaces, edges or component structure. They are not suitable for design, simulation or production. Meshes derived from them are usually coarse.
  • Dimensional accuracy: The accuracy depends on the capture and the calibration. For binding dimensions, you need a laser scan or a survey.
  • Difficult surfaces: Reflective, transparent or textureless surfaces, which are typical for stainless steel and glass, cause artifacts.
  • Baked-in light: The lighting of the capture is part of the scene. Relighting it later is possible only to a limited degree.
  • Static scenes: You cannot animate moving parts. For moving people, volumetric video is more suitable.
  • Data privacy: Captures in the hall quickly show people, screens or confidential systems. Clean them up before you publish them (see data privacy).

Frequently Asked Questions (FAQ)

Which equipment do I need for the capture?

For smaller objects, a current smartphone is sufficient. For halls, system cameras with a wide-angle lens or 360-degree cameras are common, and for outdoor facilities, drones. The important factors are even light, sufficient overlap and steady camera movement.

Can I make a CAD model from a splat?

Not in a useful way. Splats describe the appearance, not the geometry. To get back to CAD, reverse engineering with a laser scanner or structured light scanner is the correct method.

Do Gaussian splats run in the browser and on headsets?

Yes. Several web engines and XR platforms support splats. But you must compress large scenes or divide them into levels of detail, so that mobile devices and standalone headsets keep running smoothly.

How large are splat files?

Uncompressed as PLY, they quickly reach several hundred megabytes to gigabytes. With compression formats such as SOG or SPZ, they shrink to a fraction of that size. It also helps to crop unnecessary areas such as the ceiling or outdoor areas.

How can you combine splats with machine models?

In a shared scene: the splat capture supplies the real environment, and the machine comes as a glTF model from the CAD data. With glTF as a shared container, this combination becomes easier and easier.

Industry Standards and Technologies

Capture: - LiDAR – laser-based distance measurement - Volumetric video – moving 3D captures - Reverse engineering – from the component to the CAD model - Photogrammetry – 3D from photos as a mesh - NeRF – neural radiance fields

Display: - Real-time rendering – display in real time - WebGL – 3D in the browser - Virtual environment – digital spaces

Applications: - Digital twin – digital replica - Virtual prototype – planning before construction - Virtual reality – immersive tours - Immersive experience – trade shows and presentations