Robust computer vision models exhibit perceptually-aligned gradients, indicating generative capabilities.
Perceptually-aligned gradients in robust models are explained through off-manifold robustness.