Nano Banana Pro redraw: a white-on-black mask as a second reference

google/nano-banana-pro has no mask_url on Sume, so Sume's own Moodboard Redraw tool edits one region another way. It sends the source image first and a white-on-black mask PNG second, both as references, and the prompt says to edit only the white region and keep the rest identical. You can copy that pattern in an API call.

What Redraw sends

When you paint a region on a moodboard tile, the tool exports your strokes as a PNG: black everywhere, white where you painted. That image is attached after the source. If nothing was painted, the tool falls back to a local-change prompt without a mask.

The prompt wording

The mask prompt names both images and the rule. It says the first image is the source, the second is a white-on-black mask, to edit only the white region, then gives your change, then asks to match the original lighting, style and identity outside the mask. Copy that order.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"google/nano-banana-pro","prompt":"Edit this image. The first image is the source. The second image is a white-on-black mask: edit only the white region; keep everything else identical. In the masked region: replace the mug with a small potted plant. Match the original lighting, style, and identity outside the mask.","aspect_ratio":"auto","input_references":[{"type":"image_url","image_url":{"url":"https://example.com/desk.png"}},{"type":"image_url","image_url":{"url":"https://example.com/desk-mask.png"}}]}'

How this differs from mask_url

A real mask_url is a field the model gets as a mask. A mask sent as a second reference is an instruction the model follows by reading a picture, so it is looser. Expect some bleed past the edge, and check the unmasked areas against the original.

When to pick which

Use mask_url on GPT Image 2.5 when the boundary matters, for example a label or a face. Use the second-reference method when you want Nano Banana Pro's look or its 4K tier, the change is loose, and a slightly soft edge is fine. Keep aspect_ratio: "auto" so the output keeps the source shape. See Nano Banana Pro aspect ratio auto.

Make the mask correctly

Match the mask size to the source, use pure white and pure black, and avoid soft gray unless you want a feathered edit. Add a small margin around the object, because the model repaints the white area but not an exact outline. Host both files at public HTTPS URLs.