Not even wrong, in the Pauli sense: to engage requires ceding the incorrect premises that image models only accept text as input and that the generation process relies on this text