Museum exhibit intros on Wan 3.0: 22 s each, eight for $22.00 at 720p

A 22-second Wan 3.0 clip costs $1.375 at 480p, $2.75 at 720p and $5.50 at 1080p on Sume. Two 11-second jobs cost exactly the same, so the choice is about continuity, not money (read 2026-10-09).

One job or two for a 22-second intro

An exhibit intro is a slow pan across a hall, a title, a close-up of an object. Wan 3.0 reaches 30 seconds in a single job, so 22 seconds fits without chaining.

Submit with a photo of the hall

Feed the object or room photo as a first frame so the camera starts on something real. frame_images entries need a frame_type of first_frame or last_frame.

import os, time, requests

API = "https://api.sume.com/v1/videos"
H = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}
body = {"model": "wan-3.0", "prompt": "Slow dolly through a quiet museum gallery, dust in the light, ending on a glass case", "duration": 22, "resolution": "720p", "aspect_ratio": "16:9", "frame_images": [{"type": "image_url", "image_url": {"url": "https://example.com/hall.jpg"}, "frame_type": "first_frame"}]}
job = requests.post(API, headers={**H, "Idempotency-Key": "museum-intro-001"}, json=body, timeout=60).json()
deadline = time.time() + 1800
while time.time() < deadline:
    time.sleep(30)
    job = requests.get(job["polling_url"], headers=H, timeout=60).json()
    if job["status"] == "completed":
        print(job["unsigned_urls"][0], job["usage"]["cost"])
        break
    if job["status"] in ("failed", "cancelled"):
        print(job["status"], job.get("error"))
        break

A hall of eight exhibits

Eight exhibit intros at 22 seconds and 720p cost 8 x $2.75 = $22.00. At 480p the set is $11.00; at 1080p it is $44.00.

If a gallery screen loops only the first 15 seconds, ordering 15 seconds saves $0.875 per clip at 720p. Order the length the screen will show rather than trimming later.

The Wan 3.0 envelope runs to 30 seconds, so a slow museum pan still has headroom: 30 seconds at 720p is $3.75.

What to check

• Splitting only helps when you want to review halfway; the second job would need its own start frame to match the first one's last frame.
• A failed job releases its reservation where applicable, so a retry with the same Idempotency-Key after a creation failure does not stack two holds.
• Keep the final text overlays (exhibit name, dates) out of the model prompt and add them in a timeline step.
• Cancel only works while the job is still queued; once generation starts the cancel call returns 409 job_generation_already_started.