Transparent PNG: generate it or cut it out?

To get a transparent PNG on Sume, send background set to transparent and output_format png to POST /v1/images with ChatGPT Image 2.5 (openai/gpt-image-2.5). If you already have the image and only need the subject, use a cutout job (rmbg_create on hosted MCP) instead.

Two routes

The image docs list ChatGPT Image 2.5 with up to 16 image references, an optional mask_url and the background option. output_format takes png, jpeg, webp or svg. The same page says that for transparent stills today you can also use Image 1.0 with transparency true, so check which route fits your integration.

Generate directly

Set output_format to png, since the other formats listed are jpeg, webp and svg and you want an alpha channel. POST /v1/images blocks up to 30 seconds and returns 200 with a hosted URL, or 202 with a job envelope if it runs longer; check the status code, not the body shape.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "A single brass key, studio lit, centered, nothing else in frame",
    "background": "transparent",
    "output_format": "png",
    "aspect_ratio": "1:1"
  }'

Which route to pick

Generate directly when the subject does not exist yet. The model draws it with no backdrop to remove, so you skip a second paid job. Cut out when the subject is already a photo you must keep as is, because regenerating it would change it.

Whichever you choose, put the result on a checker or a dark and a light background before you ship it, since faint halos show up only against a contrasting color.

• Generated: background transparent, output_format png.
• Existing photo: a cutout job, then check the edges.
• Always download the file and test it on dark and light backgrounds.