What is Gaussian splatting? A plain guide to 3DGS for cars
Gaussian splatting, often shortened to 3DGS (3D Gaussian Splatting), is one of the fastest-moving techniques in 3D graphics. In a few years it went from a research paper to something you can run on a phone video of your own car. This guide explains what Gaussian splatting is, how it differs from traditional 3D models, why it works so well for cars, and how you can try it yourself without building a pipeline.
Gaussian splatting in one sentence
Gaussian splatting represents a scene as millions of tiny, soft, coloured blobs ("Gaussians") floating in 3D space. Each blob has a position, a size, a rotation, a colour and a transparency. When you look at the scene from a given angle, all those blobs are projected ("splatted") onto your screen and blended together. The result looks remarkably close to a photo, but you can move around it freely.
How is that different from a normal 3D model?
Most 3D models you know, from games to car configurators, are meshes: surfaces made of triangles with textures painted on top. Meshes are great when an artist builds an object by hand, or when you need clean geometry for engineering. Creating a realistic mesh of a real car from photos, however, is hard. Reflections, thin parts, glass and fine details often turn into blobs or holes.
Gaussian splatting skips the surface entirely. It does not try to decide exactly where the paint ends and the air begins. Instead, it learns how the car *looks* from every angle it has seen. That is why splats are so good at:
- Reflections and gloss that change as you move around
- Fine detail like grilles, badges, tyre sidewalls and stitching
- Soft edges and transparent parts where meshes usually struggle
- Photorealism without hours of manual texturing
The trade-off is that a splat is not a clean, editable CAD surface. It is a visual copy. For viewing, presenting, sharing and creative work, that is often exactly what you need.
How does a Gaussian splat get made?
The 3DGS process roughly follows four steps:
1. Capture. You record many images of the object from different angles. A steady video works well, because it gives hundreds of overlapping frames. 2. Camera reconstruction. Software works out where the camera was for every frame. This is called structure-from-motion and produces a sparse point cloud as a starting point. 3. Training. Starting from those points, an optimiser places, grows, splits and removes Gaussians until rendered views match the original frames as closely as possible. Tens of thousands of iterations are common. 4. Export. The final set of Gaussians is saved, usually as a .ply file, which you can open in viewers such as SuperSplat or in game engines with a 3DGS plugin.
Each step has pitfalls. Blurry frames, bad camera poses, a messy background or the ground under the car can all hurt the final result. That is why a general-purpose tool does not always give great results on cars.
Why Gaussian splatting suits cars
Cars are an ideal but demanding subject for 3DGS. They are large, you can walk all the way around them, and their appeal is mostly visual: paint, shape, light. At the same time they are glossy, partly reflective and often parked on a busy street or in a cramped workshop.
A car-specific 3DGS pipeline can make choices a generic tool cannot. It can pick the sharpest frames from your video, reconstruct the camera path around a vehicle, recognise the ground and straighten the model so it stands level, and cut away irrelevant background. That is exactly what CarTwin does.
What can you use a car splat for?
People use Gaussian splatting of cars for very different goals:
- Car businesses and photographers who want an interactive 3D view next to their photos
- 3D and CGI artists who need a realistic reference or a background vehicle
- Game and real-time developers experimenting with 3DGS assets in Unreal Engine or Unity
- AR/VR studios prototyping experiences with real cars
- Automotive tech teams and researchers exploring neural rendering and 3D datasets
- Enthusiasts who simply want a 3D copy of their own car to share
How CarTwin turns one phone video into a 3DGS car
CarTwin is a Gaussian-splatting service built specifically for cars. You do not need a camera rig, a GPU or any knowledge of training parameters. The workflow is simple:
1. Film the car with your phone. Walk two calm rounds around the car: one overview round with the whole car in view, and one detail round where you move the camera slowly up and down. Our recording guide shows exactly how. 2. Upload the video. Upload the file as it is, from your phone or computer. Uploads can resume if your connection drops. 3. Wait roughly 60 to 75 minutes. Our pipeline selects frames, reconstructs the cameras, trains the Gaussian splat and aligns the car with the ground. 4. View, share and download. Open the result in the 3D viewer in your browser, share a link, or download the .ply file to use in SuperSplat, a game engine or your own tools.
One scan costs one credit. You can find the current bundles on the pricing page.
Honest expectations
CarTwin is still fully in development and in an early pilot phase. A splat is a visual copy, not a perfect one-to-one measurement of the car, and the outcome depends heavily on the video. Shiny dark paint, harsh sunlight or a car that is only partly in frame can lead to artefacts. If a scan fails because of an error in our software, you get the credit back. A disappointing result on its own does not lead to a refund, so it pays to follow the recording tips carefully.
Frequently asked questions
Is Gaussian splatting the same as photogrammetry?
No. Both start from photos, but photogrammetry builds a mesh with textures, while Gaussian splatting builds a cloud of soft blobs optimised for realistic rendering. We compare them in detail in 3DGS vs. photogrammetry for cars.
Do I need special equipment?
No. A recent smartphone that films in good quality is enough. Steady movement and good light matter more than the camera.
What file do I get?
A .ply file containing the Gaussian splat. Read how to open your .ply in SuperSplat for the next steps.
Try Gaussian splatting on your own car
The best way to understand 3DGS is to see your own car as a splat. Have a look at the examples, read the recording guide, and then start your first scan. One phone video is all it takes.