Recast a video longer than 30 seconds: trim, recast, rejoin

To swap the people in a video longer than 30 seconds, split it into ranges of 5 to 30 seconds, run H3 Max Recast on each range with the same reference photos, and join the results. On Sume that is a video trim per range, one Recast job per range, and a Timeline plan to put the pieces back together.

Sume does not split or stitch inside the Recast call. The row accepts one source of 5 to 30 seconds and the duration you send is the source's length. The steps below are a manual pipeline built from documented Sume surfaces, and the one step I could not confirm from the docs, how a joined render treats each clip's own audio, is flagged where it comes up.

What is the limit, and why 30 seconds?

Sume's Video Router docs give h3-max-recast 5 to 30 seconds, with the duration equal to the source length. Sume's API reference adds that no single shot in the source may run over 15 seconds. fal's H3 Max Recast page (read 2026-10-02) lists 768p at $0.30 per second and 1080p at $0.45 per second of output, and does not state a maximum duration, so the 30 second cap here is Sume's documented contract rather than something I could confirm on the vendor page.

That shot rule shapes where you cut. A 25 second range that is one continuous take is inside the duration limit and outside the shot rule.

How do you plan the ranges?

Start by inspecting the source. Video inspect returns probe facts for one clip on media.sume.com, with an optional stills grid, and probe plus stills are unbilled. Cut on shot changes, not on a fixed grid, so each range starts and ends on a cut and no person is swapped mid-motion across a seam.

Keep every range between 5 and 30 seconds. A tail shorter than 5 seconds is refused by Recast, so merge a short tail into the previous range, or extend the previous range's end, instead of leaving it alone.

What does the trim step look like?

Video trim takes start plus exactly one of end or duration, and returns a new MP4 artifact. precision: exact is the default and re-encodes frame-accurately, which is what you want for cuts that must land on a shot boundary. The job costs $0.02 per trim according to the docs, which are the source for that number.

Run one trim per range, each with its own Idempotency-Key, then submit each trimmed clip to Recast with the inspected length of that clip as duration.

curl -X POST https://api.sume.com/v1/video-trim \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: recast-part-1" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/long.mp4",
    "start": 0,
    "duration": 30
  }'

How do you put the pieces back?

Timeline renders a video list over a separate audio spine: audio.url plus audio.duration_seconds, and up to 200 video[] slots whose URLs must already live on media.sume.com. Its plan route is unbilled and tells you the compiled duration and cost before you render.

The open question is audio. Fal says Recast preserves the source audio, but Timeline's documented model is one audio spine you provide, so the safe route is to detach the original video's audio once as wav, use it as the spine, and run the unbilled plan to check lengths line up. If a recast range is a different length than its source, the plan is where you find out, not the paid render.

Also expect small continuity differences between ranges: each Recast job is independent, and Sume does not promise that the same person looks identical across separate jobs. Reusing the same photos is the only control you have.