Trained with ~200 images. Works remarkably well. Not much else to say about it, honestly.