Image Editing API: Edits & Inpainting with One Key

Most "generate an image" APIs stop at text-to-image — editing an existing photo (retouch a background, swap an object, inpaint a masked region) usually means a different provider, a different SDK, or both. On Aigenway, image_edit is just another method on the same provider/model/method contract: one source image, one instruction prompt, one API key, whether the model behind it is Google's Nano Banana or OpenAI's gpt-image-1.

Two editing models, one method name

Both google_banana and openai_image expose image_edit alongside their text_to_image method, so switching providers mid-project is a one-line change, not a new integration. Confirm the exact model codes live on your account before you build against either — providers ship and retire versions on their own schedule:

curl "https://api.aigenway.com/v1/methods?provider=google_banana" \
  -H "Authorization: Bearer sk_live_..."

curl "https://api.aigenway.com/v1/methods?provider=openai_image" \
  -H "Authorization: Bearer sk_live_..."

Each response lists the current model codes, which methods they support, and per-image pricing on your account right now.

Edit with an instruction prompt (Nano Banana)

Nano Banana's image_edit takes a single source image plus a plain-language instruction — no mask required. Pass the image URL in input_images and describe exactly what should change and what should stay the same:

curl -X POST https://api.aigenway.com/v1/generations \
  -H "Authorization: Bearer sk_live_..." \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "google_banana",
    "model": "gemini-3.1-flash-image-preview",
    "method": "image_edit",
    "params": {
      "prompt": "remove the parked car in the background, keep everything else unchanged",
      "input_images": ["https://example.com/street-photo.jpg"]
    }
  }'

Nano Banana is built around a single reference image per request — it does not accept a separate mask, so the prompt itself has to carry the scope of the edit ("only the background," "keep the subject's pose and lighting").

Masked inpainting (gpt-image-1)

gpt-image-1 supports the same image_edit method but adds an optional PNG mask, which is what makes it an inpainting tool rather than just an instruction-follower — the transparent region of the mask marks exactly where the model is allowed to draw, and everything outside it is left untouched pixel-for-pixel:

curl -X POST https://api.aigenway.com/v1/generations \
  -H "Authorization: Bearer sk_live_..." \
  -H "Idempotency-Key: $(uuidgen)" \
  -H "Content-Type: application/json" \
  -d '{
    "provider": "openai_image",
    "model": "gpt-image-1",
    "method": "image_edit",
    "params": {
      "prompt": "fill the masked area with a wooden bookshelf",
      "input_images": ["https://example.com/living-room.png"],
      "mask": "https://example.com/living-room-mask.png",
      "resolution": "1024x1024"
    }
  }'

The mask must match the source image's dimensions, with alpha transparency marking the editable region. Without a mask, gpt-image-1 still edits the whole image from the prompt alone, same as Nano Banana — the mask is what buys you a hard-edged, region-locked result.

Picking a model for the job

• No mask, fast iteration — Nano Banana's instruction-only image_edit is the quicker loop when the change is describable in a sentence ("swap the shirt color to navy").
• Precise region control — gpt-image-1's mask support is the better fit for product photos, UI mockups, or any edit where the untouched pixels have to stay bit-identical.
• Both return a task, not an inline image. Poll GET /v1/generations/{task_id} or set a callback_url to get a webhook the moment the edit finishes — the task shape is identical across every model on the gateway, so the polling or webhook code you write once works for both. Full contract in the task lifecycle and webhooks docs.

Why one gateway beats a provider-only account

• No separate GCP or OpenAI account. No per-provider API key to provision, no separate invoice to reconcile — sign up once and call the endpoint.
• One integration, every editing model. The same provider/model/method/params shape that edits with Nano Banana also drives gpt-image-1, and every future editing model added to the gateway.
• Real USD billing. A prepaid balance replaces per-provider invoicing; check what's left from the API at any time.
• Idempotency and webhooks included. Retry-safe requests and signed callbacks are already built into the task contract, not something you bolt on per provider.

Pull the current model codes from /v1/methods, pick instruction-only or masked editing based on how precise the change needs to be, and ship — the request shape doesn't change either way.