









In an internal Vivid 3D study, fine-tuning a COCO-pretrained detector on 1,600 synthetic renders raised AP50:95 on real car photos from 80.42% to 84.87%. The test was small: 34 real photos with 43 annotated cars. Performance on the broader COCO car benchmark fell, which is why we treat this as evidence for task-specific adaptation and not a general uplift. Read the synthetic data case study for the setup and limits.

