Sora's short-clip dialogue rule, tested on Gemini Omni Flash

OpenAI's Sora 2 Prompting Guide says a 4 second clip fits one or two short exchanges of dialogue, and that shorter clips follow instructions more reliably. Carry the same limit to Gemini Omni Flash 1.1, which outputs 3 to 10 seconds with native audio always on, and verify the spoken words with a transcript rather than your ears.

Google's Omni docs (read 2026-10-03) say English is fully supported and other languages were not evaluated, and that voice editing is unsupported. Treat the Sora rule as a starting point for Omni, not a documented Omni limit.

A budget of lines per clip

The Sora guide's example is the useful part: very short clips hold very little speech. A practical translation is one short line per 3 to 4 seconds, with the line written out in quotes and the speaker described. Omni has no published words-per-second limit, so keep lines short and test.

Verify with a transcript

Video inspect can return a transcript for a Sume-hosted clip. Compare it with your script line by line. A missing last word usually means the line did not fit; shorten it or lengthen the clip.

Request

Quote the words and name the speaker in the prompt.

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: dialogue-budget-001" \
  -d '{
    "model": "gemini-omni-flash-1.1",
    "prompt": "A barista at a counter says, \"Your coffee is ready.\" The customer smiles and replies, \"Thank you.\" Single continuous shot, warm light.",
    "resolution": "720p",
    "duration": 5,
    "aspect_ratio": "16:9",
    "mode": "async"
  }'

What to do on a miss

Do not resubmit a paid request blindly. Change one thing (shorter line, longer clip) and run again. Existing posts cover Seedance and Kling transcripts in the same way.