Qwen-Image 2.1 native RGBA vs Sume's ChatGPT Image 2.5 transparency

Qwen-Image 2.1's Hugging Face card (read 2026-10-05) lists native RGBA transparency, meaning the model outputs an alpha channel itself. Sume does not host Qwen-Image 2.1. Its transparent route today is background: "transparent" on ChatGPT Image 2.5, which the Sume docs say is the only model that accepts the field. The card names the Qwen Research License Agreement, so read it before you use the weights commercially.

Transparency routes

Compare where the alpha comes from.

What native alpha changes

A model that outputs RGBA directly skips the cut-out step, so there is no segmentation pass that can trim fine edges like hair or glass. That is a design claim from the card's feature list, not a measured result, so run your own edge test on a hard subject. On Sume, a post-hoc cut-out is a second call with its own cost and its own edge behaviour.

Licence check

The model card names the Qwen Research License Agreement. We did not read the full licence text for this post, so we state no terms. If you plan commercial use, read it or ask the publisher. Hosted rows on Sume are separate models and a separate service; see Qwen-Image 2.1's non-commercial licence against hosted Qwen on Sume.

A transparent request on Sume

Ask for PNG with a transparent background. If the reply is 200, the URL in data[0].url is the image.

import asyncio, os, httpx

async def main():
    body = {
        "model": "openai/gpt-image-2.5",
        "prompt": "a glass perfume bottle, studio product cut-out",
        "background": "transparent",
        "output_format": "png",
    }
    headers = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}
    async with httpx.AsyncClient(timeout=60) as c:
        r = await c.post("https://api.sume.com/v1/images", headers=headers, json=body)
    j = r.json()
    print(r.status_code, j["data"][0]["url"] if r.status_code == 200 else j)

asyncio.run(main())

Practical advice

If you only need transparent stills and no local GPU, the Sume route with background: "transparent" is one call. If you already run the weights and your use is research, native RGBA saves a step. In both cases keep a copy of the alpha-bearing PNG: converting to JPEG removes the transparency, and Sume's output_format allows png, jpeg and webp.