FLUX Erase: mask rules and dilate_pixels vs a Sume mask edit

FLUX Erase is a dedicated object-removal endpoint: you send an image and a black and white mask, white marks what to remove, and dilate_pixels (0 to 25, default 10) grows the mask so edges are covered. Sume has no erase endpoint; the nearest route is a masked edit with mask_url and a prompt on ChatGPT Image 2.5.

BFL facts are from its FLUX Erase page and May 21, 2026 release note; Sume facts from the Image API docs, read 2026-09-30.

What are the FLUX Erase mask rules?

The endpoint is POST /v1/flux-tools/erase-v1. The mask is a black and white image where white (255) means remove and black (0) means keep, and it must match the input image dimensions. No prompt is sent; the server applies a fixed erase instruction internally. If a visible halo remains around the removed object, BFL says to raise dilate_pixels; for soft edges such as hair or smoke it suggests 15 to 20.

What does Sume offer instead?

ChatGPT Image 2.5 (openai/gpt-image-2.5) takes references, an optional mask_url, and background. The mask is a public HTTPS URL. Other models reject it: the code throws an unsupported-parameter error for mask_url on any other model. Image 1.0 has its own field, mask_image_url, used with image_urls.

Because the edit is prompted, you describe the result yourself, for example what should fill the area. There is no dilate_pixels equivalent in the docs, so grow the mask in your own tool before uploading it.

What should I match on a Sume edit?

On edit and image-to-image calls the docs say to prefer aspect_ratio: "auto" so the result matches the reference; omitting the field is not the same as auto. For field details across both masked routes, see mask_url vs mask_image_url and the `mask_url` how-to.

Can I get a prompt-free erase on Sume?

Not as a documented mode. Treat it as a prompted edit and check the result around the mask edge.