Trained on 2000 images and is pretty flexible. works from 0.5-1 but best between 0.8 and 1