Practice — Text-to-Image Generation with Diffusion Models (6 questions)

Your team is deciding between an autoregressive image-tokenizer-based approach (like the one in autoregressive-image-generation) and a diffusion-based approach for a new consumer text-to-image product targeting broad, open-ended prompts ("anything a user might describe"). A teammate argues: "autoregressive is simpler to build and train, and we already have that architecture working from an earlier project — let's just add text conditioning to it instead of taking on diffusion's complexity."

1. Name the three axes interviewers expect you to compare the two approaches on, and state where each approach wins.
2. Given this product's requirements (broad domain, quality-sensitive, "near real-time" rather than millisecond latency), argue for the choice you'd actually make.
3. What single sampling-time property of diffusion is the strongest argument for it in this specific product, and why does the autoregressive approach not have an equivalent?