I think you answered it yourself, stick it into a multimodal LLM.