← All articles

Why we built a car-specific Gaussian splatting pipeline

There are plenty of tools that turn video into a Gaussian splat. So why did we build CarTwin, a Gaussian splatting pipeline specifically for cars? Because cars are a surprisingly difficult subject, and a general-purpose 3DGS tool leaves most of that difficulty to you. In this article we explain the problems we ran into, the choices we made, and what that means for anyone who wants a 3DGS car model without building a pipeline.

The promise and the reality of 3DGS

Gaussian splatting can produce stunningly realistic 3D captures from ordinary video. If you are new to the technique, start with what is Gaussian splatting? In demos, the results look effortless. In practice, a typical 3DGS workflow involves:

  • Extracting frames from the video
  • Running structure-from-motion to reconstruct camera poses
  • Training the splat with dozens of parameters
  • Cleaning floaters, background and ground
  • Aligning and exporting the result

Each step can fail silently. On a car, many of them do.

Why cars are hard

Glossy, reflective surfaces

Car paint reflects the world around it. Reflections move as the camera moves, which confuses both camera reconstruction and training. We wrote a full article on this: why shiny paint is hard for Gaussian splatting.

Large smooth panels with little texture

Doors, bonnets and roofs offer few distinctive features to match between frames. That makes camera poses less reliable.

The ground matters

A car stands on something. Generic tools often produce a model that is tilted, floating or merged with a blotchy patch of tarmac. For viewing, a car should stand level on a sensible ground.

Busy backgrounds

Cars are filmed on streets, in car parks and in workshops. Background clutter adds millions of Gaussians that have nothing to do with the car and makes files heavy.

Real-world video

People film with phones, at walking pace, in imperfect light. A pipeline has to cope with some blur and unevenness rather than expecting perfect studio footage.

What our pipeline does differently

We designed every step around one subject: a car, filmed with a phone, in two rounds.

1. Smart frame selection

Instead of using every frame, the pipeline selects a fixed set of the sharpest, best-spread frames from your video. Blurry frames are skipped. This gives the camera reconstruction cleaner input and keeps processing time predictable.

2. Camera reconstruction around a vehicle

The capture pattern is known: a circle around the car, then a second round with the camera moving up and down. That knowledge helps reconstruct a consistent camera path around the vehicle, which is the foundation for everything that follows.

3. Training tuned for cars

The Gaussian splat is trained with settings chosen for car-sized subjects and long training runs, so fine details like grilles, lights and wheels have a chance to come out.

4. Ground detection and straightening

The pipeline detects the ground plane and works out which way is up. The car is straightened so it stands level, and the viewer uses that information to keep the camera above the ground and orbit naturally around the car, like a turntable.

5. Background trimming

Parts of the scene that clearly do not belong to the car are cut away, which makes the model lighter and cleaner to view and edit.

6. Compressed output

The result is delivered as a compressed splat of about 70 MB (typically 65 to 75 MB), which loads in a browser and opens directly in tools like SuperSplat.

What that means for you

You do not need to know anything about structure-from-motion, iteration counts or point-cloud cleaning. The workflow is:

1. Film two calm rounds around the car with your phone. See how to film a car for 3DGS. 2. Upload the video as it is. Uploads resume if your connection drops. 3. Wait roughly 60 to 75 minutes while the pipeline does its work. 4. View, share and download. Orbit around the car in the browser viewer, send a share link, or download the .ply for SuperSplat, Unreal Engine or your own tools.

Who it is for

We built CarTwin for two groups. First, car businesses, garages and photographers who want an interactive 3D view of a car. Second, people who want 3DGS car assets without building a pipeline themselves: 3D and CGI artists, game artists, AR/VR studios, automotive tech teams, researchers and developers.

Where we are today

We want to be open about this: CarTwin is still fully in development and in an early pilot phase. The result is a visual 3D copy of the car, and it will not always be a perfect one-to-one copy. Very glossy paint, harsh sunlight or incomplete footage still cause artefacts. If a scan fails because of an error in our software, you get the credit back. A disappointing result on its own is not refunded, so following the recording guide makes a real difference.

We keep improving the pipeline with every scan, and early users help us shape where it goes.

Try the pipeline on your own car

The best way to see what a car-specific Gaussian splatting pipeline can do is to try it. Browse the examples, check the pricing (one scan costs one credit), and start your first scan with a single phone video.