Trained on 25 images (5 repeats, batch size 2) for 800 steps.