“The Gato model can output words and actions, but was not designed to output images.”