So presumably the model could output tokens that represent images as well? For multi-modal training data, e.g. HTML pages or PDFs, does the training data inter

So presumably the model could output tokens that represent images as well? For multi-modal training data, e.g. HTML pages or PDFs, does the training data interleave the image tokens amongst the text tokens in the same document? Slightly limited, as doesn't get juxtaposition to text in complex ways, just linear placement of images.