Seedream 4.5 edit with n=4 returns one image on Sume

If you send n: 4 with reference images to bytedance-seed/seedream-4.5 on Sume, the provider payload is built with one image, not four. Edits on that model return a single result per call, so for four takes you send four requests (or change the model). Text-to-image calls without references are not affected.

What the adapter does

In the Sume image adapter, a Seedream 4.5 request that carries images is limited to 10 reference images by the API and has the image count set to 1. The catalog still lists n as a range up to 4, which is why this is easy to miss: the parameter is valid and does not error. The 10-image limit is checked before the provider call, so an eleventh reference is a 400 rather than a silent trim.

Read the response, not the request

Count data in the response and read usage.cost for the billed amount (Jobs and results). Do not compute the cost from the n you sent.

import os, requests

r = requests.post(
    "https://api.sume.com/v1/images",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={
        "model": "bytedance-seed/seedream-4.5",
        "prompt": "Same sofa, olive green fabric",
        "input_references": [{"image_url": {"url": "https://example.com/sofa.jpg"}}],
        "n": 4,
    },
    timeout=120,
)
body = r.json()
print(r.status_code, len(body.get("data", [])), body.get("usage"))

What to do instead

• Send four calls with four different prompts, which is also better for variants because each can name a different colour.
• Run them from a small pool and respect your plan's queue limit, as described in Generation admission.
• Choose a model whose edit path honours the count if you want one-call variants. Check the live GET /v1/images/models entry first.

When this does not matter

If you only ever want one edited image per call, nothing changes for you. It matters when a UI offers four variants by default and you reuse the same code path for text-to-image and edits, because the edit path will quietly show one tile instead of four.

Build the variant grid yourself

For a four-up grid, loop over four prompts, collect the four URLs and show them together. Give each call its own short label, such as the colour name, so the grid is readable and so a chosen tile can be traced back to its prompt.

Keep the same input reference in all four calls. Then the only difference between tiles is the prompt, which is what you want when a buyer is comparing finishes or colours on one product.