Understand the car first.
Then change the background.
Most "background replacements" cut out a silhouette and drop in a picture. We first recognize the make, model and real dimensions — and only then build a scene where the car actually stands correctly.
Four steps to a finished photo
Detection and identification
Boxes: car, wheels, plate. DINOv3 Car ID returns the make, model, body type and five dimensions.
Segmentation
Masks for the body, glass, wheels and plate. The body itself is never repainted — only the light around it.
Scene and light
The background is built for the known dimensions and angle: floor plane, perspective, contact shadow, reflections in the glass.
Plate and assembly
Your plate is fitted into the perspective and light. The result is 2048 px, 4:3, with the source’s fine detail intact.
Our own model.
Millions of photos.
A fine-tuned DINOv3 encoder with two heads: make-and-model classification over a broad catalog of makes and models, and regression of five dimensions in millimeters. Three checkpoints vote by median, so an error on a rare model doesn’t skew the size.
Runs on our servers.
Or on yours.
This is how our cloud runs — about 600 frames an hour. We set up the same system on your own servers, too, if everything needs to stay on your own network.