Seedance 2.5 takes 10 video references; Omni 1.1 limits them to 3 s

ByteDance Seed lists up to 10 video clips per Seedance 2.5 pass but, on the page I read (2026-10-05), does not state a per-clip length. Google says Gemini Omni 1.1 Flash video references are up to 3 seconds (read 2026-10-05), and Sume's docs for gemini-omni-flash-1.1 list up to 3 video references, each 3 seconds or less. So the practical rule on Sume: if your source clips are longer than 3 seconds, Omni needs them trimmed; for Seedance, read the catalog entry and test one longer clip.

The comparison

Trim to the move you need

For Omni, cut each reference to 3 seconds or less with video trim. It takes video_url (a clip on media.sume.com), start and exactly one of end or duration, and returns a new MP4 for $0.02 per job. Use precision: "exact" (the default) so the cut is frame-accurate; keyframe can start a GOP early.

curl -X POST https://api.sume.com/v1/video-trim \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: ref-trim-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/long-move.mp4",
    "start": 4,
    "duration": 3
  }'

Pick the clip's best 3 seconds

• Choose the segment where the motion you want is clearest, not the opening.
• Avoid cuts inside the reference; one continuous move reads better.
• Remove burned-in captions and logos first with a crop from video filter.
• Keep one idea per reference; ten different moves in one prompt blur each other.

Common failures with video references

A reference clip with a hard cut inside it often confuses the result, because the model has to guess which shot you mean. A clip with a watermark may reproduce the watermark. A clip with its own audio may leak sound you did not want. Check each of these on a still and a short listen before you attach it, and trim or crop to remove the problem.

• Cut inside the reference: trim to one shot.
• Watermark: crop it out first.
• Unwanted audio: check has_audio with video inspect.
• Wrong length: for Omni, 3 seconds or less.

Which model for which reference set

If you have a library of ten short clips that show a look, Seedance is the model that advertises room for all of them. If you have one three-second move and a product photo, Omni 1.1 on Sume is a fit, with the tags <IMAGE_REF_0> and <VIDEO_REF_0> to say what each does. Do not force a ten-clip brief into Omni; split it into several jobs instead. Group the clips by the move they show, run one job per group, and join the results on the timeline. That also gives you smaller, cheaper retries when one group misses.

Once the clips are ready, reference them in order. Omni uses <VIDEO_REF_0>-style tags per the Sume docs. Sources: Seed, Google.