When should an LLM write the image as code, and when should you call a diffusion model? Three decisions, one failure-mode asymmetry, and measured costs from Amazon Nova to Fable on Amazon Bedrock.
Code or diffusion? A field guide to programmatic image generation