Probe a clip before using it as a Seedance 2.5 reference

Seedance 2.5 accepts up to 10 video references and 10 audio references per pass, according to ByteDance Seed (read 2026-10-05). Every reference you attach shapes the result, so check each source clip first. Sume's video inspect probes one hosted clip and returns the probe facts, sampled stills and an optional transcript, so you can see length, whether it has audio, and whether anyone speaks before you attach it.

One call per clip

POST /v1/video-inspect needs video_url (a media.sume.com artifact in your workspace, imported with POST /v1/media-imports) and an Idempotency-Key. The default mode is sync: it waits up to 30 seconds for a 200, then falls back to 202 for you to poll. With nothing else sent you get 8 mid-bin stills. Send frames: false for a probe only.

curl -X POST https://api.sume.com/v1/video-inspect \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: inspect-ref-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/ref-clip.mp4",
    "frames": { "fps": 1, "seek": "fast" }
  }'

What to read in the result

Reading the results quickly

Look at the stills as a contact sheet. If six of eight frames show the same thing, the clip has less to offer than its length suggests, and a trimmed version will do. If the subject only appears in the last third, trim to that range.

Keep the inspect ids with the project. A week later, the record of what each reference contained is worth more than the few cents it cost.

• Contact-sheet the stills.
• Trim to the part that matters.
• Store inspect ids with job ids.

When to skip it

For a clip you filmed yourself a minute ago and already watched, a probe adds little. It earns its cost on inherited footage, downloads, clips from several people and anything long. It also helps when a job came back with unexpected sound: the probe tells you which input carried audio. Pair it with a transcript when the clip has speech: at $0.01 per audio minute it is a cheap way to confirm exactly what words a reference would bring with it.

Limits to remember

Reject clips with branding you do not own, text you do not want echoed, and faces you do not have rights to. Then pass the remaining clips as references and keep the filenames in a list next to the job ids.

• Source up to 1800 seconds and up to 24 stills per call.
• A silent clip with transcribe: true fails with inspect_source_has_no_audio; probe has_audio first.
• Stills use a default max_edge of 768; for an exact frame at source size use video frames instead.
• Inspect bills probe and stills by their own compute; the transcript adds its per-minute rate.