Transparent product cutout, GPT Image 2.5: webp alpha, 7 cents

To get a product cutout with a transparent background from ChatGPT Image 2.5, send background: "transparent" and an output_format of png or webp to POST /v1/images on Sume. At high quality it costs 7 cents per image; at low it is 1 cent. The Sume docs list background: auto|transparent|opaque as a ChatGPT Image 2.5 field, and OpenAI's guide says transparency needs png or webp.

The request

{
  "model": "openai/gpt-image-2.5",
  "prompt": "a matte black thermos bottle, centered, soft shadow-free studio look",
  "background": "transparent",
  "output_format": "webp",
  "aspect_ratio": "1:1",
  "quality": "high"
}

Why webp or png

Transparency needs an alpha channel, and jpeg has none. PNG is the safe, lossless choice. WebP also carries alpha and files are smaller, which matters for a catalog of thousands of cutouts. OpenAI's guide says png is the default format and lists jpeg and webp as options; for a transparent background it says to use png or webp.

Sume's docs list output_compression (0 to 100 for webp and jpeg) in the schema but state that no model advertises it in v1, so a request that sends it returns 400 unsupported_parameter. If you need a lighter webp, re-encode on your side.

Check the alpha before you ship

Model-made transparency can leave a faint halo or a stray background patch around fine edges. Open a few results on a dark and a light background, since a halo that is invisible on white shows on black. If a piece fails, edit it with a mask_url instead of regenerating the whole thing.

The docs also keep Image 1.0 as the older compatibility route for transparent stills with transparency: true, but it is retiring, so build new flows on POST /v1/images.

Cost for a catalog

A 500-item catalog at high quality is 500 x 7 cents = $35.00. At low quality for drafts it is $5.00. Run the low pass first to check that the prompt template gives a centered object with room around it, then run high on the approved list.

If the shot starts from a supplier photo instead of a prompt, send it in input_references (up to 16 on this row) and keep background: "transparent"; an edit adds estimated input image tokens to the output charge, so it costs a little more than the text-only call.

Placing the cutout on a page

A transparent webp or png drops onto any background colour in CSS or in a design tool. Give the object some margin in the prompt ('centered with generous empty space around it') so the soft edges are not clipped by the frame. Keep the source resolution: transparent assets scaled down look fine, and scaled up they show every edge flaw. If a marketplace asks for a pure white background with no alpha, composite the cutout on #FFFFFF in your own code rather than asking the model for a white backdrop, which can leave a faint grey tone.