Raw dealer photographs go in; catalogue frames come out. The pipeline segments the car with SAM3, replaces the background, corrects the body geometry and lens distortion so the car and its new room agree, relights the paint, restores the contact shadow, and returns a 2048-pixel frame. Thirty frames of one vehicle have to agree with each other, which is the constraint that shapes every decision in it.
Built on FLUX.2 Klein 9B with LoRA adapters trained for each stage: background replacement, geometry restoration with the room held fixed, and its mirror - geometry distortion, used to manufacture the training pairs. The model cards on Hugging Face document what each one does and where it fails.
Anastasiia Butova - ML engineer, Belgrade, Serbia. Email black.design@me.com · LinkedIn · GitHub · Telegram · Hugging Face