GPT Image 2.5 mask_url on Sume: build the PNG mask with Pillow

On Sume, mask_url is accepted only by the GPT Image 2.5 rows, openai/gpt-image-2.5 and openai/gpt-image-2.5-sunburst, and it has to be a public HTTPS URL. Other image rows return a 400 for it. The mask file itself follows OpenAI's rules: its guide (read 2026-10-06) says the mask needs an alpha channel, the same format and size as the image, and a size under 50MB.

So the work is in building the file correctly and hosting it where Sume can fetch it.

Build the mask

This script makes an opaque black PNG the same size as your source and cuts a transparent rectangle where the edit should happen. Which area the model edits is defined in OpenAI's guide, so read that page and confirm on a low-quality test call before you trust the direction.

import sys
from PIL import Image, ImageDraw

def make_mask(src: str, dst: str, box: tuple) -> None:
    size = Image.open(src).size
    mask = Image.new("RGBA", size, (0, 0, 0, 255))
    ImageDraw.Draw(mask).rectangle(box, fill=(0, 0, 0, 0))
    mask.save(dst, format="PNG")
    print("mask", size, "alpha at box start:", mask.getpixel(box[:2])[3])

if __name__ == "__main__":
    if len(sys.argv) != 7:
        raise SystemExit("usage: mask.py SRC DST X0 Y0 X1 Y1")
    x0, y0, x1, y1 = map(int, sys.argv[3:])
    make_mask(sys.argv[1], sys.argv[2], (x0, y0, x1, y1))

Three masks, three conventions

Send it

• Upload the source and the mask to storage that serves public HTTPS, with no login and no localhost.
• Send the source as the first input_references entry and the mask as mask_url.
• Use quality: "low" while you check the direction and the edge.
• If the call fails with a message about an unfetchable image, the usual cause is a link that is not public.

Do not reuse an Ideogram mask

Ideogram's convention is the opposite style: a black and white image, not an alpha cutout. A mask that works for one will not work for the other, so keep them in separate folders and name them by target model.