Clean studio look from a phone photo for a listing: edit cost by model

A reseller can send a phone photo as an input_references image and ask for a clean studio background: 200 listings cost about $4.23 on gpt-image-2.5 medium at 1024, $10.00 on Seedream 4.5 and $20.00 on Nano Banana 2.1.

eBay's own picture rules live on eBay; Sume's notes are in eBay photo requirements. Also remember that an edited image must still show the item you are selling: see eBay AI listing photos.

Cost of 200 edited listings

Prompt that keeps the item honest

Tell the model to change the surroundings only, and compare the result to the source before listing.

import os, requests

r = requests.post(
    "https://api.sume.com/v1/images",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={
        "model": "openai/gpt-image-2.5",
        "prompt": "Keep the camera body, strap and every scratch exactly as photographed. Replace only the background with a seamless light grey sweep.",
        "quality": "medium",
        "aspect_ratio": "auto",
        "input_references": [
            {"type": "image_url", "image_url": {"url": "https://example.com/camera-phone-shot.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)

Use a mask when the item must not change

gpt-image-2.5 also takes an optional mask_url (a public HTTPS URL) for edits. Mask the background to keep the edit off the item, and still check the result visually.

Quality gate

Do a cheap automated gate before upload.

• Compare the output's aspect ratio to the source; aspect_ratio: "auto" should match it.
• Show the original and the edit side by side to a person for the first batch.
• Never edit away damage or wear that a buyer needs to see.

Why gpt-image-2.5 costs more here

gpt-image-2.5 is token-priced, so an input reference and prompt text add to the bare 1024 text price ($0.007375 low, $0.0165 medium). The table already includes one reference and a prompt of about 140 characters. Each extra reference adds more input tokens, so check the live pricing line for the model before a large run.