Omni Flash 1.1 takes no audio reference: where a Short's voice goes

Gemini Omni Flash 1.1 on Sume lists image_url and video_url in supported_input_references but not audio_url, so you cannot hand it a narration file to follow. Seedance 2.x, Wan 3.0 and the MiniMax H3 models do accept audio references. For a voiced Omni Short, either let the model generate synced audio, or put your voice on the Timeline audio spine.

A 60-second vertical Short from six 10 s Omni clips at 720p is $7.50, plus $0.10 for the Timeline render: $7.60. Shorts run up to 3 minutes (YouTube Help, read 2026-10-09).

Who takes what

Docs list which models accept which reference types; the catalog repeats them per model.

Two ways to a voiced Omni Short

Option A is native audio: Omni returns a synced track by default, and generate_audio is the switch. Option B is a spine: generate silent or native clips, then render with your voice file as audio.url. In Timeline the spine sets the output length, and a soundtrack can sit under it with duck_db.

curl -X POST https://api.sume.com/v1/videos \
  -H "Authorization: Bearer $SUME_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"gemini-omni-flash-1.1","prompt":"Scene 2: ...",
   "aspect_ratio":"9:16","resolution":"720p","duration":10,
   "generate_audio":false}'
# generate_audio false -> narration comes from the Timeline spine

Gotchas

• Sending an audio_url reference to Omni is a request the model's catalog entry does not list; check supported_input_references before you build around it.
• Omni takes 3 to 10 seconds a job, so a 60-second Short is at least six jobs.
• Omni's video edit mode goes through the Video Router video_url field and is separate from generation.
• A Timeline spine needs audio.duration_seconds from 1 to 1800 and a Sume-hosted file.