Reference remix of a clip you own: what Sume's first read returns

Reference ingest is the read that comes before a remix. Sume reference ingest takes one media.sume.com clip your workspace already owns, up to 300 seconds, and answers a manifest with shots, on-screen text, audio facts and a labeled strip. It is read-only: the source is never re-encoded and no MP4 is produced.

What is in the manifest

• shots[]: frame-exact cuts that tile the clip, with one sharp keyframe per shot.
• text_tracks[]: Korean and Latin on-screen text read at source resolution. Low-confidence lines are marked needs_verification with a crop.
• audio: a loudness gate, speech detection, beats when the track is music, and a transcript when it transcribes.
• overview: one strip of up to six tiles with timecode labels.

Cost, honestly

There is no flat price. Every read reserves its Modal compute ceiling and captures its own container seconds times the Modal list rate times 1.25, plus the platform fee, never above the hold. For reference_remix, speech transcription is on by default and adds the per-minute transcript rate to the reservation; Sume settles to what ran and settles to zero when the track is silent.

Errors that name the fix

Use the audio fact

If audio.silent is true, the docs say to plan new music and discard the source audio, and never request transcription on it. That is the cue to generate a track with the Music Router rather than hunt for the original sound.

Request

curl -X POST https://api.sume.com/v1/reference-ingest \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: ingest-1" \
  -d '{"video_url":"https://media.sume.com/artifacts/artf_demo/reference.mp4","purpose":"reference_remix"}'

The default mode is sync: it waits up to 30 seconds and returns 200 with the manifest, or 202 with a queued job to poll.