It's supposed to use a hash of the image, so they technically need the exact same image. Doesn't seem like they're using machine learning.