tl;dr: Image generation models fall into two camps: diffusion models (cheap, fast, seed-controllable, but literal) and multimodal models like GPT Image and Gemini (better at instructions, text, and editing, but pricier and less controllable). You pay per canvas regardless of complexity, prices range 100x across providers, and effective prompting means describing the production process (camera, film, print) rather than hype words. Build thin provider abstractions since APIs change constantly, offload exact work (sizing, text, transparency) to deterministic code, and use agents to write prompts, critique outputs, and iterate—while keeping humans in charge of taste.