Yes, it could just make one call to a multimodal llm to describe the scene