Trained on 40 images from 1920. works good between 0,5 and 1