Perhaps freshen it up and extend the test by feeding the model the rendered output so it can iterate once. Assuming a multi-modal model.