AI image generator with a transparent background: 2 steps
Make an AI image with a transparent background in two steps: generate it on a plain backdrop, then remove the background to get a PNG with alpha.

To get an AI image with a transparent background, generate the subject on a plain, contrasting background, then run background removal on the result. The cutout is a PNG with an alpha channel, so everything around the subject is see-through and the graphic sits cleanly on any page, slide or photo. With the Sume API that is two calls, POST /v1/images and then POST /v1/rmbg-1.0/remove with the new image's URL; the transparency comes from the second call.
Generation facts below come from Sume's Image API docs. Background removal (RMBG 1.0) and upscaling are specified in the OpenAPI document behind the Sume API reference. Both were read on 2026-09-28.
How do I prompt for an image that cuts out cleanly?
Background removal has to find the edge between the subject and everything else, so give it an easy edge. These are prompt habits, not settings:
- One subject, whole and centered, with space around it. Anything cropped by the frame edge ends in a straight cut.
- A plain background in a color the subject doesn't use: light gray behind a dark product, dark gray behind a pale one.
- No cast shadow, floor, reflection or props, so nothing but the subject sits on the background.
- Several candidates at once.
nsets how many images one call returns; the current catalog allows up to 4 on ChatGPT Image 2.5 (openai/gpt-image-2.5). Cut out only the one you keep, since every cutout is a separate billed job.
How do I generate the image and cut it out with the API?
Two requests, the second fed by the first:
POST /v1/imageswithmodelandprompt. It waits up to 30 seconds. A200carries the images indata[].url; a202carries a job to poll atGET /v1/jobs/{id}/statusand read atGET /v1/jobs/{id}/result.POST /v1/rmbg-1.0/removewith the chosen image's URL asimage_url, the only required field. The schema asks for a public HTTPS URL and prefers a Sume media URL, which is what the first call returns.- The cutout runs as a job,
asyncby default. Poll its status until it iscompleted,failedorcanceled; a completed job's result holds PNG artifacts with alpha, mirrored to Sume. - Download the files you keep.
/v1/imagesreturns signed URLs.
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 ceramic teapot, whole object centered, plain light gray background, soft even light, no shadow",
"n": 4,
"output_format": "png"
}'
curl -X POST https://api.sume.com/v1/rmbg-1.0/remove \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{ "image_url": "https://media.sume.com/img/01J.../0.png" }'Should I upscale before or after removing the background?
Before. POST /v1/image-upscale-1.0/upscale enlarges an image by an upscale_factor from 1 to 4 (default 2) and returns a PNG by default, or a JPG; in current code a webp request comes back as PNG. The API reference doesn't describe how the upscaler treats an alpha channel, so enlarge the opaque image and cut out the larger file. The cutout's price does not vary by image size. AI image upscaler API covers the upscale call.
How much does a transparent AI image cost?
Two charges. The image is priced per model, and ChatGPT Image 2.5 estimates each request's price from its size and quality; in current code its catalog price line is for one 1024 × 1024 image at high. So read the real figure from the response: usage.cost is the USD amount billed. Failed or canceled generations are not billed. AI image generator API cost compares the models.
The cutout is listed at $0.0225 per image on API pricing, plus a 5.5% agent fee by default.
What are the limits?
Each step is its own request and returns new files. For an image you already have, Remove background API covers the cutout call on its own.
- One image per cutout job, with no format or color option: the result is a PNG with alpha.
- Keep the cutout as a PNG. JPEG has no alpha channel, so a JPG copy fills the transparent area with a solid color; Can a JPG have a transparent background? explains why.
- Check the edges at full size before you publish, especially around hair, glass and smoke.
| Step | Endpoint | Takes | Returns |
|---|---|---|---|
| Generate | POST /v1/images | model and prompt; optional n, output_format | PNG, JPEG or WebP at signed, Sume-hosted URLs |
| Upscale (optional) | POST /v1/image-upscale-1.0/upscale | Public HTTPS image_url, upscale_factor 1–4 | PNG (default) or JPG |
| Cut out | POST /v1/rmbg-1.0/remove | One public HTTPS image_url | PNG with alpha |
Sources
Related posts
More in Media tools
- Amazon Sponsored Display video specs: display ads with video
Amazon display ads with video, formerly Sponsored Display, take 16:9 at 1920×1080 or more, 6–45 seconds, up to 500 MB, and autoplay muted.
- Amazon Sponsored Products video: specs and guidelines
Sponsored Products video takes 1–5 silent feature videos per product: 7 seconds or longer, 16:9 (1920×1080 recommended), MP4 or MOV up to 500 MB.
- Cinematic bars size: bar height for 2.39:1 and 2.35:1
Cinematic bars on a 1920×1080 video are 138 px each for 2.39:1 and 131 px for 2.35:1. The formula, sizes for 720p to 4K, and how to add them.
- How to combine photos and videos into one video
Combine photos and videos into one video on a single timeline: each photo held for a set time, each clip cut to length, one soundtrack under it all.
Written by Sume