Solo dev here. I built Nano Banana to make image generation feel instant: typical time-to-image ~1–2s at 1024px.
What’s live: text-to-image, image-to-image with reference uploads, login + small free trial.
Latency/infra: p50 ≈ 1.2s, p95 ≈ 2.0–2.1s under moderate load; cost ≈ 0.7–2¢/image (settings dependent).
How it’s fast: warm workers, tiny FIFO queue (1 job/user when busy), prompt-hash caching, early streaming preview, one fast retry; rate limits + basic content filters.
Would love feedback on the architecture, better ways to measure “time-to-image,” and pricing (subs vs credit packs, small team plan).
reply