That's because it's a two-way street, a multi-modal model that is highly proficient at real-life image generation is also highly proficient at interpreting real-life image input, which is something sorely needed for robotics.