I'm unsure why you're using an LLM to generate images. Don't we already have models (some made by the same company) that do this?