I wonder if it's actually leaking training data and these are real images, as per this: https://theaiunderwriter.substack.com/p/an-image-of-an-arche...