2026-10-27 · AI in production
Cutting out vehicle photos with AI: the model was the smaller part
A vehicle dealer has the images from its 360° studio processed by an external service. The question was whether we could do it ourselves at least as well. Segmentation was solved quickly — the difficulty lay elsewhere.
The studio photographs vehicles on a turntable: a rotation of many high-resolution frames plus stills. The service cuts the vehicle out, adds a new background, creates shadow and reflection and crops. The specification said only “remove background, white, margins”. What actually happens was written nowhere.
First the recipe, then the model
Before writing a line of pipeline I compared real pairs of original and result, image by image and pixel by pixel. That gave the actual recipe: one scale for the whole rotation, paint colour untouched, a light-grey background with a horizon line, a soft shadow plus a dark zone under the floor pan, a blurred reflection. For one kind of still the image was only darkened evenly — there was practically no processing at all.
For the cut-out, two AI models independently recommended BiRefNet, MIT-licensed; we ruled out models with closed commercial weights.
Where the work really was
The ground line is not horizontal. Draw it straight and the vehicle “floats” because of perspective. The right line is the lower convex hull of the silhouette — the points where the wheels touch down. The reflection is mirrored column by column around it, and the shadow under the floor sits only between the wheels.
I measured with a blind test: a model acting as judge did not know which image came from whom. The first version lost 4:6; after three fixes to shadow, glass and edge it won 7:5. One of the shadow fixes made the floor lighter instead of darker — caught only by a measured brightness profile, not by eye.
Humans are stricter than the judge
Looking at the side-by-side, the client still rated the existing service slightly better. The analysis showed two things no specification mentions: the service replaces whatever is visible through the windows with its own tinted background, which makes the interior look “cleaner”. And for small vehicles it does not enlarge the vehicle but shrinks the canvas; our upscaling softened the image.
On the other hand our version is sharper and keeps fine details such as antennas that the service cuts off. Segmentation takes about two seconds per image on an Apple M2 Max. As of today this is a prototype; the client decided to take it to production.
What I take away: before replacing an AI service, reconstruct its recipe from inputs and outputs. Part of the “processing” is often nearly empty, and quality is decided by geometry and light, not by the model.