Swap a product background with qwen-image: 2.5 cents, one reference

qwen-image costs $0.025 per image on Sume and accepts up to 10 input_references, so a background swap is one request with the product photo attached: 1,000 swaps are $25.00. Its larger sibling qwen-image-max is $0.0938 but is listed with zero references, which means text-to-image only.

The request

Describe the new scene and say what to keep. aspect_ratio: "auto" keeps the source shape.

import os, requests

r = requests.post(
    "https://api.sume.com/v1/images",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={
        "model": "qwen/qwen-image",
        "prompt": "Keep the sneaker exactly as is. Place it on a wet concrete street at dusk.",
        "aspect_ratio": "auto",
        "input_references": [
            {"type": "image_url", "image_url": {"url": "https://example.com/sneaker.jpg"}}
        ],
    },
    timeout=60,
)
print(r.status_code)
body = r.json()
if r.status_code == 200:
    for item in body["data"]:
        print(item["url"])
else:
    print(body)

qwen-image vs qwen-image-max

When it is not enough

A background swap without a mask can alter the product itself. If the product must stay pixel-exact, cut it out first with RMBG and composite it onto a generated scene in your own code, or use gpt-image-2.5 with mask_url.

Check the output against the original in a script before it goes live; a cheap per-image rate makes it affordable to regenerate rejects.

A safer two-step

Cut the product out with RMBG ($0.0225 per image), generate only the scene, and composite the two in code. The product is then your pixels, not a model's redraw.

• Total per image: $0.0475 for RMBG plus one qwen-image scene.
• One scene can serve many products if the lighting direction matches.
• Read GET /v1/images/models/{model_id}/endpoints for the live pricing line before a large run: prices are catalog data and can change.