Nano Banana 2.1 inpainting: Google's prompt template, no mask on Sume

Nano Banana 2.1 does not take a mask image. Google's guide calls its method "inpainting (semantic masking)" and does it with a sentence: name the one element to change and tell the model to keep everything else the same. On Sume that is the only way for this model, because mask_url works on the GPT Image 2.5 rows alone.

What is Google's template?

The guide gives this template, which you fill in with the element and its replacement:

• Using the provided image, change only the [specific element] to [new element/description].
• Keep everything else in the image exactly the same, preserving the original style, lighting, and composition.

What does the Sume call look like?

Send the image as an input_references item and the template as the prompt. Use aspect_ratio: "auto" so the result matches the reference's shape; if you omit it, the result is not the same as auto. Reference URLs must be public HTTPS.

{
  "model": "google/nano-banana-2.1",
  "prompt": "Using the provided image of a living room, change only the blue sofa to a vintage brown leather chesterfield. Keep everything else unchanged.",
  "input_references": [
    {
      "type": "image_url",
      "image_url": {
        "url": "https://example.com/living-room.jpg"
      }
    }
  ],
  "aspect_ratio": "auto",
  "resolution": "2K"
}

When should I use a mask instead?

Use mask_url when the region is hard to name or you need it to be exact. The GPT Image 2.5 rows (Flare and Sunburst) accept mask_url and up to 16 references. OpenAI's guide notes that masking with GPT Image is entirely prompt-based, so the shape may not be followed with complete precision (read 2026-10-09). The mask is guidance, not a hard clip, on either route.

Sending mask_url to Nano Banana 2.1 returns 400 unsupported_parameter. See mask_url on GPT Image 2.5 for that route.

What does each route cost?

Nano Banana 2.1 is $0.10, $0.15 or $0.20 billed per image by tier. GPT Image 2.5 depends on quality and size; read the endpoint line in the image models docs before you batch.

Whatever route you choose, compare the result to the source image at 100 percent. Prompt-only edits can drift lighting or texture outside the named region, and a mask is guidance for GPT Image, not a hard clip. A quick difference blend in your editor shows exactly which pixels changed, and it takes less time than another paid call.