Restaurant promo clip: keep Gemini Omni's native audio or swap it

A restaurant promo clip from gemini-omni-flash-1.1 always comes with synced audio, because Sume's docs reject generate_audio: false on that model. Keep the sizzle and room tone when they help, and when they do not, drop the audio with Video trim (audio: "drop", $0.02) and put a music bed under it in Timeline.

Which to choose depends on the shot. Sources: Video Router and Video trim, read 2026-10-04.

Three ways to handle the sound

The three routes differ in cost and in how much control you have over the sound.

Prompt for the sound you want

For a plated-dish close-up, ask for the sound in the prompt: sizzle, a knife on a board, a soft clink. For an exterior or a street shot, expect traffic and voices you may not want; trim the audio away.

curl -X POST https://api.sume.com/v1/video-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: ramen-clip-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "Close-up of a bowl of ramen being finished with chili oil. Steam, a spoon dipping in. Only kitchen sound: broth bubbling, a ladle tap. No speech, no music.",
    "image_url": "https://example.com/ramen.jpg",
    "resolution": "1080p",
    "aspect_ratio": "9:16",
    "duration": 6,
    "mode": "async"
  }'

Check for speech

Native speech is the risk. A model may add voices over a dining-room shot. Read the clip before you publish, and if you caption it, remember that speech-only captions cover speech and not sound effects.

Replace it

To replace the sound, trim with audio: "drop" and send the silent MP4 into a one-slot Timeline render with a soundtrack of your own bed. fade_out_seconds on the soundtrack must fit inside the output length, and duck_db needs a real spine (not silence), so for a silent-picture promo skip ducking.