Recolor a product with a swatch reference and mask on GPT Image 2.5

BFL's FLUX 3 Image documentation lists three edit operations: recolor, replace and move, targeted with bounding boxes written as [y0, x0, y1, x1] on a 0 to 1000 scale. A catalog request for a recolor is common in product work: same product, new color.

Sume does not list FLUX 3 and does not take bounding boxes. Its region control is mask_url, accepted only on ChatGPT Image 2.5 (both variants), with up to 16 references.

Mapping the job

A request

OpenAI's guide says the image to edit and the mask must be the same format and size, and the mask needs an alpha channel, so export it at the source dimensions.

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":"Recolor only the masked jacket to the swatch color in the second image; keep fabric texture and shadows","quality":"high","mask_url":"https://example.com/jacket-mask.png","input_references":[{"type":"image_url","image_url":{"url":"https://example.com/jacket.png"}},{"type":"image_url","image_url":{"url":"https://example.com/swatch.png"}}]}'

Limits of the swap

A mask restricts the region but a model can still shift tone nearby. Check edges and shadows on the first few outputs, and rerun with a tighter mask if the color bleeds. Converting a bounding box to a mask is a few lines of Pillow: scale 0 to 1000 to pixel coordinates and fill the rectangle.

How this was checked

Vendor facts come from the pages listed in the sources, read on 2026-10-05. Sume facts come from the Image API docs and the catalog code on main on the same date. Catalogs and limits change, so read the descriptors from GET /v1/images/models before you pin a number in production code.