Replace the music on a recast clip: Timeline spine, $0.10 render

H3 Max Recast swaps the people in a clip and leaves the rest alone. fal's page says it preserves the source "motion, camera, cuts, and audio", and its schema returns the video "with the source sound" (both read 2026-10-04). That is the right behavior for a voiced testimonial and the wrong one when the source carries a track you cannot license or a bed that clashes with your brand.

To change the sound after the swap, put the recast clip into Timeline 1.0 and declare the audio yourself. Timeline takes one audio spine plus ordered video slots, so the sound of the render is the spine you name, and the clip supplies the picture.

The render call

Every URL on Timeline must already be a media.sume.com artifact or asset in your workspace, so import the recast clip and your music file first. Then send audio.url with audio.duration_seconds equal to the clip length and one video[] slot starting at 0.

Timeline's default output is 1080 by 1920. A landscape source needs output.width and output.height set, or it is fitted into a vertical frame. The example below is a 12 second landscape clip.

curl -X POST https://api.sume.com/v1/timeline-1.0/render \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: recast-new-bed-001" \
  -d '{
    "audio": {
      "url": "https://media.sume.com/artifacts/artf_demo/new-bed.wav",
      "duration_seconds": 12
    },
    "video": [
      {
        "source_url": "https://media.sume.com/artifacts/artf_demo/recast.mp4",
        "start": 0,
        "duration": 12
      }
    ],
    "output": { "width": 1920, "height": 1080 }
  }'

What it costs and what to check

Timeline's public rate is $0.10 per output minute, rounded up, and the job uses no provider inference, only worker ffmpeg. A 12 second render therefore reserves one minute: $0.10 on top of the Recast clip.

Two checks keep the swap honest. First, the spine must cover the clip: a slot can end at most 0.5 seconds before the spine does. Second, listen to the result, or run audio detach on the render and compare the new file's duration_seconds with the clip, so you know the old track is gone.

If the clip had speech you want to keep, a new bed alone will cut it. In that case record the line again with a text-to-speech job and use that as the spine, then lay the music in as soundtrack with duck_db so the voice stays on top.