Migrating to Unified Image API from Google AI Studio, fal.ai or Replicate

Moving an existing Nano Banana integration here is usually an afternoon: the concepts map one-to-one, and the differences are the parts you'll be glad to delete. Below are the mappings we're asked about most.

The mental model#

From Google AI Studio#

Google's API is synchronous and returns the image inline as base64. Two changes:

1. Split the call. POST /v1/generations (get id) then poll GET /v1/generations/{id} or use webhook_url. Your timeout problems disappear: the submit returns in milliseconds regardless of model load.
2. Fetch instead of decode. Results come as presigned URLs valid 15 minutes — stream them to your storage instead of holding base64 blobs in memory.

- POST https://generativelanguage.googleapis.com/v1beta/models/gemini-2.5-flash-image:generateContent
- {"contents":[{"parts":[{"text":"a red fox in snow"}]}]}
+ POST $BASE/v1/generations
+ {"model":"nanobanana","prompt":"a red fox in snow","resolution":"1K","ratio":"1:1"}

The prompt string itself transfers unchanged — same underlying model family, same prompting behaviour.

From Replicate#

The shape is already familiar (create prediction → poll/webhook). What changes:

• version hashes → stable model ids (nanobanana, nanobanana-pro) listed at GET /v1/models; no pinning ceremony, routing and failover happen behind the id.
• input: {...} → flat body: prompt, resolution (1K/2K/4K), ratio, n, seed, plus images for image-to-image. The legacy size field ("1024x1024") is still accepted and mapped to the nearest tier and ratio.
• Billing: per-second GPU pricing → fixed credits per image, known before you submit (cost in the response), refunded on failure.

The checklist#

1. Create a key, put real traffic's IP ranges in the key's allowlist.
2. Wrap submit with an Idempotency-Key per user intent (see our idempotency post).
3. Map your error handling to code values — the errors table is short.
4. Run both providers in parallel for a day (credits are prepaid; a test run on the Starter pack costs a few dollars) and compare failure rates before switching DNS-style.

Full request/response schemas with real payloads: API reference. Per-model pricing in USD: models.