Wan 3.0: one 30-second prompt or three 10-second clips on Sume

On Sume, one 30 second Wan 3.0 request with three scenes in the prompt and three 10 second requests cost the same, $3.75 at 720p, so the choice is about control. Three clips let you redo one weak scene for $1.25; one request is simpler and keeps a single continuous take.

What the vendor claims

Alibaba's Wan 3.0 repository lists "auto scene splitting" and multi-person and human-object interactions among its features (read 2026-10-05). It does not, in the part we read, tell you how the model decides where a cut falls. The Sume docs describe Wan 3.0 as a 2 to 30 second text, image and reference model; scene splitting is not a request field on Sume, so you describe the scenes in the prompt.

Same total, different risk

The price is per output second, so how you slice the 30 seconds does not change the total. What changes is the cost of a mistake.

What the redo saves

With one weak scene in three, splitting saves $2.50 at 720p. If all three scenes work the first time, the two plans cost the same and the single request needs less joining.

When one request is right

A single request, written as time ranges, is a good fit when the scenes are one continuous idea, for example a product on a table with the camera moving through three angles. Three requests fit when scenes differ in location, subject or style, and when a client approves each one.

When three clips are right

Use the same style sentence in each of the three prompts, and say in each what the previous and next scene are so the motion lines up. If you need the end of scene one to match the start of scene two, take the last frame of the first clip and send it as the first frame of the next with image_url. That keeps continuity without any extra model.

Tracking jobs

Both approaches need a stable way to track jobs. Use Idempotency-Key values like film-7-scene-1-v1 so a retry cannot double bill a scene, and poll or use a webhook as described in the Sume jobs docs; a 30 second job should not use a synchronous wait.

{
  "model": "wan-3.0",
  "prompt": "0-10 s: dawn over a harbor, slow pan. 10-20 s: a fisher mends a net, close-up. 20-30 s: boats leave the harbor, wide shot. Same soft, golden style throughout.",
  "resolution": "720p",
  "duration": 30,
  "aspect_ratio": "16:9"
}

Review and bookkeeping

The review step differs too. A single 30 second clip needs a full watch to find a problem, and the problem may be at 22 seconds. Three clips give you three short reviews and a place to stop early: if scene one is wrong at 480p, you have spent $0.625 and not $1.875.

Idempotency and ledgers are simpler with one job and more useful with three. With one request, a single job id covers the film. With three, keep a small table of scene number, key, job id and status, so you can resume after a crash without resubmitting a scene you already paid for.

Parallel scenes

Concurrency favors the split as well. Sume runs jobs up to your plan limit (1 on Free, 4 on Pro, 8 on Startup, 20 on Scale) and queues the rest. Three 10 second jobs can run in parallel on Pro, while a single 30 second job runs alone. Run time per clip is not documented in the pages we read, so we make no promise about total time, only that parallel jobs can overlap.

Check where the cuts land

Time-range prompts are a request for structure, not a guarantee. Check where the cuts actually fall in the result, and if they land wrong, switch to three clips. The Sume range is 2 to 30 seconds per request; sources are the Wan 3.0 repository (read 2026-10-05) and the Video Router docs.