I used a mix of on-set close-ups and other photos in the dataset.