Why is everyone pretending it's the LLM that is creating the image and not the diffusion model?