They probably didn't train it on 3d renders. Training it on 3d renders would likely hurt its ability to handle real photos.