[version_five]

Exactly, because it's not actually probing attribution, it's just finding the most similar images in the training data. You can just go to https://rom1504.github.io/clip-retrieval/?back=https%3A%2F%2... and do this yourself without the hyperbole