Twelve colourways times three takes: n=3 versus 36 calls on Sume

For 12 colourways with 3 takes each, send 12 requests with n: 3, one per colourway, instead of 36 requests with n: 1. The price is identical, because Sume bills cost_usd times n, but you make a third of the calls and get the three takes of one colourway together.

Each colourway still needs its own prompt, since the colour is in the words. The n field repeats one prompt; it does not vary it.

Cost of the batch

Thirty-six images at the catalog base prices cost the same however you slice them. Models and base prices were read on 2026-10-10; resolution and quality settings can raise a request above the base price.

The request, per colourway

Send the same product photo as input_references[0] in all twelve calls, change only the colour phrase in the prompt, and set n to 3. Every model in the table allows n of 1 to 4, so 3 is within the range. Keep the prompt otherwise identical, so the differences you see between colourways are colour.

{
  "model": "black-forest-labs/flux.2-pro",
  "prompt": "The same sneaker recoloured in sage green with cream laces, plain white studio background",
  "n": 3,
  "aspect_ratio": "1:1",
  "input_references": [
    { "type": "image_url", "image_url": { "url": "https://example.com/sneaker-base.png" } }
  ]
}

Why fewer requests helps

Twelve requests are easier to track than thirty-six. Each can return 200 with the images or 202 with a job, so there are twelve places to handle the slow path instead of thirty-six, and a smaller burst is less likely to meet 429 queue_full, which means the workspace cannot take another paid job. The Errors and rate limits page says to back off and use retry-after when it is present.

A failure also costs less to reason about. Sume does not bill a failed or cancelled generation, so if one of the twelve calls fails, you re-send only that colourway.

Review and selection

Three takes of one colourway are easier to judge side by side than three takes scattered across a batch. Put the twelve responses into a contact sheet or a simple grid in your own tooling, one row per colourway, and choose the best of each row. Because the takes share a prompt, the differences you see are the model's variation, not your wording.

If a row has no usable take, rerun that colourway alone with n: 3 and a small prompt change. The cost of that rerun at the base price is 3 x 0.0375 = 0.1125 USD on FLUX.2 Pro, which is easy to absorb in a 1.35 USD batch.

When to split anyway

Use single-image calls if the colourways need different references, as when each has a different base photo, or if you want to review each take before you pay for the next. Also use them on Grok Image, whose n range is 1 to 1 in the catalog, so it cannot do this at all.

• Name the colour with a plain phrase and add the hex value only if you also check the result.
• Review the three takes together and keep one per colourway.
• Log the prompt text next to each URL so a colourway can be re-run exactly.