Gaussian splatting is a new approach to representing three-dimensional environments. Instead of approximating surfaces with polygons, many small, semi-transparent "splats" are placed in the space. Each splat has a position, a color and an extension. Together, they can depict fine structures, soft shadows and transparent materials so that photorealistic 3D scenes are displayed in real time. This beginner's guide explains the basics of the process, shows you how to capture high-quality data and which tools you need. The focus keyword Gaussian Splatting Beginner's Guide is used to ensure good findability in search engines.
Recording the material - Create photos or videos from different perspectives.
Alignment - All images are located in relation to each other in space ("sparse point cloud").
Training - The software generates the splats from the images.
Cleaning & Optimization - Remove interference points and reduce file size.
Export & Presentation - Provide the model on the web or in VR/AR.
In the following section, we explain each of these steps in detail.
A good recording is the basis for a convincing splats model. Many errors in data recording are difficult to correct later. You should therefore pay attention to the following aspects.
First of all, the scene should be as static as possible. People, animals or blowing trees create ghost images or holes in the reconstruction. Choose a lighting situation without extreme contrasts and avoid glare spots in the lens. Switch off auto exposure and auto white balance on your camera and select constant settings instead. A low ISO value produces little image noise, while a small aperture (around f/8-f/11) ensures sharpness throughout. Use short exposure times (1/125 s or shorter) to prevent motion blur.
For a robust 3D reconstruction, it is more important to take many different angles than simply taking many photos at the same position. Walk around your object and vary the height to create a strong parallax effect. The turntable method has proven itself with objects: Take photos from even distances in several rings. A grid of positions covering walls and corners is suitable for rooms or buildings. Make sure there is sufficient overlap (70-80 %) between neighboring photos and leave an angle difference of 15-30 degrees between the positions. It is better to take 100-300 high-quality photos than thousands of blurred images. For videos: record at least 4K at 60 fps and extract the sharpest frames later.
The choice of focal length has a significant influence on the result. A wide-angle lens (around 24 mm full format) offers plenty of context and provides a strong perspective. You should avoid ultrawide lenses under 20 mm because they create distortions that make alignment more difficult. A slightly longer lens can be used for long distances, such as drone shots. If possible, shoot in RAW format and keep the focal length constant during a session; with zoom lenses, you should not change the zoom. If you are working with a smartphone, make sure that features such as HDR and night mode are switched off in order to obtain consistent exposures.
The first step in the workflow is called Structure-from-Motion (SfM). Here, the exact positions and viewing directions of the camera are determined from many overlapping photos or from a video. This process is also known as alignment because all images are geometrically located in space. Alignment works in the same way as in classic photogrammetry: distinctive image points (features) are recognized between the images and the position of the camera is calculated by triangulation. For this reason, a photogrammetric point cloud can usually be generated from images created for Gaussian splatting without any problems - conversely, you can use image material created for photogrammetry to train splats. The difference between the two methods lies mainly in the following steps: For Gaussian splatting, millions of splats are optimized from the point cloud, while photogrammetry creates a triangulated mesh. In addition, Gaussian splatting often requires more images from different angles in order to train the volumetric effects in a stable manner.
Training is the heart of the process. Here, millions of splats are optimized from the camera positions and images until they are as close as possible to the original photos.
Depending on the amount of data and hardware, this step takes several hours - especially if several resolution levels are calculated.
Postshot is currently the most accessible tool for beginners: it performs alignment, pose optimization and training in one interface. Experienced users can also work with scripts or refine the parameters manually.
After training, there are almost always small imperfections - "floaters" or noise artifacts. These are removed in post-processing and the file is compressed.
SuperSplat enables manual selection and deletion of splats in the browser.
SplatTransform CLI offers automatic filters, e.g. according to position, transparency or brightness.
The goal is a clean, high-performance file that is suitable for real-time applications.
Once the splat model has been cleaned up, it can be exported and shared online.
With the SuperSplat library or one WebGL engine (PlayCanvas, A-Frame), the model can be rendered directly in the browser - ideal for portfolio pages, museum projects or WebAR experiences.
Whether smartphone, action cam or DSLR - every system has strengths and weaknesses. Below you will find an overview of which camera type is suitable for which area of application.
| Camera type | Target group and use | Advantages | Restrictions |
|---|---|---|---|
| Smartphone | Beginners and spontaneous projects | Always with you, easy to use, modern smartphones deliver surprisingly good results thanks to computational photography | Fixed focal length, small sensor, many photos are needed for large scenes |
| Action camera | Outdoor enthusiasts, sports photography | Robust housing, wide angle, integrated stabilization | Small sensor and little manual control, distortions due to fisheye optics |
| DSLR or mirrorless camera | Professional projects, large buildings or architecture | Large sensors, interchangeable lenses, full control over exposure and focus | Higher weight, requires experience in handling manual settings |
Start with the device you are most familiar with. A modern smartphone may be sufficient for small objects or quick tests. For ambitious projects, however, we recommend a DSLR or mirrorless camera with a suitable wide-angle lens. If you regularly want to capture large buildings or interiors, it is worth purchasing a dedicated 3D scanner.
For extensive scenes such as halls, castles or entire city districts, a portable 3D scanner is an enormous relief. The PortalCam combines four 12-megapixel cameras with a high-resolution LiDAR sensor to create a dense point cloud in a short space of time. The device weighs less than one kilogram, scans around 200 square meters in 15 minutes and offers a high level of accuracy thanks to its integrated AI. Such scanners are ideal for construction and architectural documentation, for film productions or for virtual tours. However, you must be prepared to invest in professional hardware and proprietary software. If you prefer to process the data yourself, use models that export open formats such as PLY.
Once you have recorded the data, it needs to be processed. Several programs and services support the complete workflow from alignment to export of the splats.
Postshot is a commercial desktop application that is particularly aimed at beginners. It imports your video or image data, automatically calculates the camera poses and optimizes the splats. Postshot selects the best photos on request and offers several quality profiles with which you can control the level of detail. The advantage lies in the simple user interface and the integrated shooting guidelines. After completing the training, you can export the model as a browser viewer or share it as a compressed file. You can find further information on the official website of Postshot
There are several freely available options for users who would like to tinker or explore scientific models. COLMAP is a proven structure-from-motion pipeline; it calculates precise camera poses and generates the initial point cloud. nerfstudio is a flexible framework for the development of neural radiance fields and splats models; it can be controlled via scripts and offers numerous customization options. Brush in turn brings the training into the browser and uses WebGPU, so you don't bog down your computer with heavy calculations.
Often also Polycam or Luma AI in connection with Gaussian Splatting. These apps are commercial cloud services. They run on mobile devices, but are limited to smaller scenes - not least because large data sets can neither be recorded with smartphone cameras in sufficient quality nor transferred efficiently to the cloud.
Polycam and Luma AI are therefore useful for gaining initial experience with the technology or scanning small objects, playing in the professional area hardly play a role. Their results are clearly limited in quality and are neither suitable for precise architectural visualization nor for productive VR/AR pipelines
A number of software solutions that implement Gaussian splatting almost automatically have now become established in the professional sector. Particularly noteworthy is the XGRIDS package, Lixel Cyber Color offers a very compact toolset. These software solutions make it possible to create 3D Gaussian splats in different resolutions and share them on the Internet without in-depth technical knowledge. Operation is highly automated, making them ideal for production workflows where efficiency and reproducibility are more important than experimental control.
The system handles camera alignment, pose reconstruction and the actual splat training fully automatically and can export different quality levels depending on the target platform - from quick preview models to high-resolution splats for film and VR projects.
The XGRIDS LCC softwarewhich was specially developed for the company's own PortalCam, falls into this category: it combines LiDAR and multi-camera data and generates a 3DGS model from it.
Even after careful recording and training, splats often contain annoying points called "floaters", or there are simply too many splats for a performant display. This is where special editors and command line tools come into play.
The SuperSplat Editor is a browser-based, free editor. It offers two view modes, Centers and Rings, with which you can view the splats from different perspectives. There are six tools to choose from, such as lasso, polygon, brush or box, with which you can precisely mark and delete unwanted splats. After editing, the splats can be scaled, rotated or merged. A practical measuring tool helps you to check real proportions. You can export finished splats as an HTML viewer and integrate them directly into websites.
The SplatTransform command line tool is a comprehensive solution for advanced users. It supports various formats such as PLY, SPLAT, KSPLAT or CSV and offers conversion and filter functions. With a single command, you can move, rotate or scale splats, remove splats outside a box or only keep splats with a certain opacity. Multiple scenes can be merged and a CSV output enables statistical analysis of the data. SplatTransform is ideal if you want to optimize large scenes or integrate them into other applications.
Gaussian splatting opens up impressive possibilities for creating realistic 3D scenes. With the Gaussian Splatting Beginner's Guide you have learned how to capture high-quality data, choose the right cameras and software and post-process your splats. Here are the most important points:
Careful preparation: Choose a static scene, constant exposure and a suitable focal length. Shake-free, sharp images are the basis for a good result.
Versatile perspectives: Move around your subject, vary the height and ensure overlapping. A few photos taken in a targeted manner deliver better results than uncontrolled mass shots.
Suitable hardware: Use smartphones for your first experiments, for larger projects use DSLR cameras or a 3D scanner. The PortalCam makes it much easier to cover large areas.
Suitable software: Postshot offers a convenient way to get started, while open source tools like COLMAP or nerfstudio give you flexibility. Don't forget to clean up your splats with an editor like SuperSplat before you publish them.
Recognize boundaries: Gaussian splatting often requires more image material than photogrammetry, but rewards you with a volumetric, detailed model. Test different methods and find your own workflow.
Are you interested in developing a virtual reality or 360° application? You may still have questions about budget and implementation. Feel free to contact me.
I am looking forward to you
Clarence Dadson CEO Design4real
