The training data includes imagery that, when interpolated over a high dimensional manifold, results in these things.

That doesn't imply that they were in the training set, or even anything close to them.