Gemini Omni 40-second extension: chain 10-second clips on Sume

Google's Gemini Omni can continue an existing clip in 10-second steps up to 40 seconds, but Sume's gemini-omni-flash-1.1 row does not expose that extend call. To get a longer scene through Sume you chain separate 10-second jobs, using the final frame of each clip as the first frame of the next.

What Google offers and what Sume exposes

Google's prompt guide says scene extension takes an existing clip and continues generating from where it left off, across multiple 10-second intervals up to 40 seconds total. Sume's Video Router doc is explicit about the other side: the API does not expose previous_interaction_id or extend, because it is not in the published input schema of fal, the provider behind the row.

The chain recipe

Generate clip one with duration: 10. Pull a frame near the end with video frames, host that frame at a public HTTPS URL, and send it as image_url on job two with a prompt that starts with the action that follows. Repeat up to four times for roughly 40 seconds. In the OpenRouter-shaped /v1/videos body the same idea is a frame_images entry with frame_type: first_frame.

• Keep style, lighting and location words identical in every prompt, so the model sees one scene.
• Say "keep everything else the same" only on edits, not on new chained shots.
• Expect small identity drift at each seam, and check frames at the cut before you stitch.

Sketch of the second job

This reuses the submit-and-poll helper from the earlier post. LAST_FRAME must be a public HTTPS image, because Sume's docs say image inputs must be reachable over public HTTPS.

import os, time, hashlib, json, requests

H = {"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"}

def run(body):
    key = "omni-" + hashlib.sha256(json.dumps(body, sort_keys=True).encode()).hexdigest()[:24]
    r = requests.post("https://api.sume.com/v1/videos", json=body,
                      headers={**H, "Idempotency-Key": key}, timeout=60)
    r.raise_for_status()
    url = r.json()["polling_url"]
    while True:
        s = requests.get(url, headers=H, timeout=30).json()
        if s["status"] in ("completed", "failed", "cancelled"):
            return s
        time.sleep(10)

s = run({"model": "gemini-omni-flash-1.1", "duration": 10, "resolution": "720p",
         "aspect_ratio": "16:9",
         "prompt": "The vendor turns and hands the fish to a child. In a single continuous shot.",
         "frame_images": [{"type": "image_url", "frame_type": "first_frame",
                           "image_url": {"url": os.environ["LAST_FRAME"]}}]})
print(s["status"], s.get("usage"))

Cost of a 40-second scene

Omni at 720p lists at $0.10 per second on Sume's catalog, billed at list times 1.25. One 10-second job is $1.00 list and $1.25 billed, so four chained jobs are $5.00 billed for 40 seconds.

Seam checks that save a re-render

A chain is also easy to resume. Because each job has its own id and a stable key, a crash after step three only repeats step four, and you never pay twice for steps one to three.

• Compare the last frame of clip one with the first frame of clip two at full size, not as a thumbnail.
• Match the lighting words word for word across prompts.
• If a character changes, add a reference image of that character to the next job.
• Keep one Idempotency-Key per chained step so a retry reuses its job.

When to use another model

If continuity matters more than one model family, Sume's catalog also lists models with longer single-shot ranges, such as seedance-2.5 at 4 to 30 seconds. Read the Video Router capabilities before you choose.