The same photograph produces different neighbours depending on whether your model was trained for objects, styles, faces or text descriptions.