Select the examples that you think come from real Imagenet1k samples after a MoCo v2 augmentation. The rest are examples from our generated datasets that maximally fool a classifier that discriminates between Imagenet1k and our generated datasets.