In the preprocessing stage, the pipeline uses SAM 2 to extract high-resolution vehicle segmentation masks and embeds the vehicle contours into a spatial-conditioning adapter to enforce geometric boundaries at the cross-attention layers.