GPT Image 2.5 inpainting with mask_url on Sume: steps and cost

To repaint only part of an image on Sume, send openai/gpt-image-2.5 a source image in input_references, a mask_url that marks the area to change, and a prompt that describes the new content. The row accepts up to 16 references and bills about $0.066 per image. mask_url and background are listed on GPT Image 2.5 only; on every other image row they return 400 unsupported_parameter.

The steps are short, and the failure cases are listed below so you spend the $0.066 once.

Steps

• Host the source image at a public URL; Sume reads references by URL.
• Make a mask image the same size as the source. Mark the area to change as the mask's editable region, following the docs for mask format.
• Host the mask and pass its URL in mask_url.
• Write the prompt for the changed area and say what must stay. Keep the ratio of the source with aspect_ratio, or use image_size for custom pixels.
• Read usage.cost on the response and save the result URL.

Request

{
  "model": "openai/gpt-image-2.5",
  "prompt": "Replace the sky with a pale evening sky; keep the rooftops unchanged",
  "input_references": [
    {"type": "image_url", "image_url": {"url": "https://example.com/street.png"}}
  ],
  "mask_url": "https://example.com/street-mask.png",
  "aspect_ratio": "3:2"
}

Limits to check

Cost of an edit session

Five mask attempts on one picture cost about $0.33. If the first masks are rough, test a single attempt before running a batch of fifty, and check that the mask URL opens in a browser with no login, since a private URL fails the same way as a bad mask. A failed generation is not billed, but a result you dislike still is.

Use background: transparent only when you need a cutout from this row. It is listed on GPT Image 2.5 and nowhere else in the image catalog.

Common failure cases

• A mask URL that needs a login: the provider cannot fetch it and the request fails.
• A mask with a different size from the source: results shift or the call is rejected.
• A prompt that describes the whole picture instead of the change: the edited area drifts in style.
• More than 16 references: the request returns a 400 from the catalog limit.

When to skip the mask

If the change is large, such as a new background for a whole product shot, a mask may add little. Send the photo as a reference with a prompt that says what to keep, and compare the result with the masked version once. The unmasked call is the same $0.066, so the test is cheap.

Use the mask when something must stay pixel-close, such as a face or a logo, and keep the masked area as small as the edit allows.

What Sume does not do

Sume does not draw the mask for you or detect the region from a text description. It also does not preview the edit before billing. The exact mask color convention is the upstream provider's; confirm it on one cheap test before you automate.