{"id":16397,"date":"2025-11-06T22:10:19","date_gmt":"2025-11-06T22:10:19","guid":{"rendered":"https:\/\/design4real.de\/?p=16397"},"modified":"2025-11-20T07:32:06","modified_gmt":"2025-11-20T07:32:06","slug":"gaussian-splatting-beginners-guide","status":"publish","type":"post","link":"https:\/\/design4real.de\/en\/gaussian-splatting-einsteiger-guide\/","title":{"rendered":"Beginner's guide to Gaussian splatting"},"content":{"rendered":"
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t\t<\/div>\n\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t<\/section>\n\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t\t
\n\t\t\t
\n\t\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t
<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

Gaussian Splatting Beginner's Guide -
Step by step to the 3D scene<\/h1>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t
<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t
<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

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<\/strong> is used to ensure good findability in search engines.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

\n\t\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t
<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t
\n\t\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"Gaussian\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"closeup\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t
\n\t\t\t
<\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t

How is a 3D model created with Gaussian splatting?<\/h2>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
\n\t\t\t\t
\n\t\t\t\t\t\t\t\t\t

Before we delve deeper, here is an overview of the process:<\/span><\/h2>

  1. Recording the material<\/strong> - Create photos or videos from different perspectives.<\/p><\/li>

  2. Alignment<\/strong> - All images are located in relation to each other in space (\"sparse point cloud\").<\/p><\/li>

  3. Training<\/strong> - The software generates the splats from the images.<\/p><\/li>

  4. Cleaning & Optimization<\/strong> - Remove interference points and reduce file size.<\/p><\/li>

  5. Export & Presentation<\/strong> - Provide the model on the web or in VR\/AR.<\/p><\/li><\/ol>

    In the following section, we explain each of these steps in detail.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

    \n\t\t\t\t
    \n\t\t\t\t\t\t\t
    \n\t\t\t
    <\/div>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t
    \n\t\t\t\n\t\t\t\t\t\t<\/span>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t\"devices_gaussian_splats\"\t\t\t\t\t\t\t\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    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.<\/p>

    Scene and recording conditions<\/h3>

    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.<\/p>

    Camera path and perspectives<\/h3>

    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.<\/p>

    Camera settings<\/h3>

    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.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

    \n\t\t\t\t
    \n\t\t\t\t\t

    2. alignment - locating the images<\/h3>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    The first step in the workflow is called Structure-from-Motion<\/em> (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.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

    \n\t\t\t\t
    \n\t\t\t\t\t\t\t
    \n\t\t\t\n\t\t\t\t\t\t<\/span>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t

    3. training - the creation of splats<\/h3>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    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.<\/p>

    Postshot<\/strong> 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.<\/p>\t\t\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t

    \n\t\t\t\t
    \n\t\t\t\t\t\t\t
    \n\t\t\t\n\t\t\t\t\t\t<\/span>\n\t\t<\/div>\n\t\t\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t

    4. cleaning and optimization<\/h3>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t

    Data acquisition: what is important<\/h3>\t\t\t\t<\/div>\n\t\t\t\t<\/div>\n\t\t\t\t
    \n\t\t\t\t
    \n\t\t\t\t\t\t\t\t\t

    After training, there are almost always small imperfections - \"floaters\" or noise artifacts. These are removed in post-processing and the file is compressed.<\/p>