Budget for 12 UGC ad variants: draft on Mini, finish on Seedance 2.5

Twelve 15-second 9:16 variants drafted on Seedance 2 Mini at 480p cost $15.84 on Sume, and finishing three winners on Seedance 2.5 at 720p adds $26.01. The whole round is $41.85.

The same twelve rendered straight on Seedance 2.5 at 720p would cost $104.04. The draft step is what keeps a variant test cheap.

The assumptions

Prices below are Sume's billed price, which is the provider list price x 1.25 rounded up to the cent, from the price book read on 2026-10-05. They are per clip. There is no seed on Sume's v1 models, so a draft cannot be re-rendered as a keeper; the keeper is a new take of the same prompt, and it may differ from the draft.

Cost by plan

Three plans for the same twelve variants.

What a draft tells you

A Mini draft at 480p answers whether the hook is clear in the first two seconds, whether the action reads, and whether the framing suits the product. It does not tell you whether skin, fabric or small text will survive at higher quality. Pick your three winners on structure, then judge detail only on the finished renders.

Do not overread a draft's motion. A tier can move a subject differently from another tier given the same prompt, so a variant that wins on Mini may need a prompt tweak on 2.5. Plan for one retake per winner in the budget: three retakes add $26.01.

Run the drafts as one batch

Submit with an Idempotency-Key per variant so a retry cannot double-bill. Replaying the same key returns the original job instead of creating a second one. The example sends one variant; loop it for twelve.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: variant-07-draft" \
  -d '{"model":"seedance-2-mini","prompt":"Handheld selfie video, a person holds up a bottle of sunscreen in bright outdoor light and says nothing, hook: squints at the sun then smiles at the bottle","duration":15,"resolution":"480p","aspect_ratio":"9:16"}'

How to pick the three winners

Score each draft on three yes-or-no questions: does the first two seconds show the product, does the action finish before the clip ends, and is there anything on screen you would have to cut. Keep drafts that get three yeses. If more than three qualify, break ties by which hook is least like the others, because a test of three near-identical ads teaches you less than a test of three different ones.

Keep the prompt text for each winner exactly, and change only one thing when you re-render on 2.5, such as adding the resolution or a first-frame photo. If you change the prompt and the resolution together, you will not know which change moved the result. Record the id of each job next to its prompt in a spreadsheet so the winner can be traced back.

Poll many jobs, not one

Twelve jobs finish at different times. Keep a list of ids, check each at a moderate interval, and download as each completes. Sume's docs support a callback_url HTTPS webhook if you would rather be told than ask, and GET /v1/jobs/{id}/status for a status check. Do not hold twelve open connections; the job is async by design.

Last, budget the retries that do not render. A failed job is the case to check against the docs before you assume you were billed; do not guess. If your finance team needs a number, quote the plan total from the table and add the one-retake line.