Qwen-Image 2.1 subject extraction vs Sume's RMBG 1.0 cutout

Qwen-Image 2.1's model card lists subject extraction from photographs and native transparent (RGBA) output. Sume does not list Qwen-Image 2.1 in the image catalog. For a product cutout today, send the photo to Sume's RMBG 1.0 route, POST /v1/rmbg-1.0/remove, which returns a PNG artifact with alpha.

What the Qwen model card says

The Hugging Face card for Qwen-Image-2.1 describes a model with 7B parameters in its visual generation component. It lists text-to-image, editing with up to 10 reference images, native RGBA generation and subject extraction. It is released under the Qwen Research License Agreement, which the card describes as proprietary rather than open source, so check the license before any commercial use.

The Sume cutout route

The OpenAPI spec describes RMBG 1.0 as a background-removal route that takes a public HTTPS image_url and returns mirrored PNG artifacts with alpha. It supports the standard mode values: async, sync, subscribe and webhook. The sync wait is capped at 30 seconds, and results are read from the job result endpoint.

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/uploads/sneaker.jpg","mode":"async"}'

Cutout or generate a transparent image

Use RMBG when you already have the photo and want the same pixels with the background removed. Use openai/gpt-image-2.5 with background: "transparent" when you want to generate a new asset on a clear background. They answer different questions: one preserves a real photograph, the other creates an illustration or render.

Limits

The Qwen model's selling points are claims from its model card, and I did not test them. If you need the open-weight model itself, run it yourself under its license. On Sume, the Qwen entries are the hosted qwen/qwen-image and qwen/qwen-image-max models, and their feature set is whatever the catalog row lists.

Which route for which job

A quick split:

• Real product photo, new background later: RMBG 1.0.
• New illustration on a clear background: GPT Image 2.5 with background: transparent.
• Open-weight experiments: the Qwen checkpoint on your own hardware, under its license.