In a follow-up
work, PerceiverIO, they generalized it to let the model also produce outputs of arbitrary size.