H3 Max Insert-Video: add a scene mid-clip on fal, and on Sume

H3 Max Insert-Video is a fal endpoint that adds a newly generated scene into an existing video at a time you choose, then returns to the original footage. Sume does not list it. On Sume you get the same result by hand: trim the original at the insert point, generate the new scene with minimax-h3-max, and join the three pieces with Timeline.

The endpoint facts are from fal's H3 Max Insert-Video page (read 2026-10-02). The Sume side is from the Video generation, Video trim and Timeline docs. The Sume docs list no Insert-Video row.

What are fal's limits and prices?

fal's page gives these numbers: a source video of up to 60 seconds and 50 MB, a start time of at least 1.625 seconds, a resume time after the start inside the 60 second limit, and a generated scene of 5 to 13 seconds at 480p or 768p. A prompt guides the scene and reference images or reference videos are optional.

Billing covers only the newly generated scene, not the whole output. That is the useful part: a 60 second source with an 8 second insert is charged for 8 seconds, plus reference tokens beyond the first 4,096.

How do I do it with Sume's own rows?

Three steps. First cut the original into a before range and an after range with video trim (start plus end or duration, $0.02 a job per the docs). Second, generate the middle with minimax-h3-max for 5 to 15 seconds. Third, assemble before, new scene, after in Timeline, whose video slots take media.sume.com URLs and whose audio is a separate spine you supply.

To make the join look continuous, generate the new scene from the last frame of the before range as a first frame. H3 Max supports first and last frame image-to-video, and a last frame taken from the start of the after range closes the gap, as described in the first and last frame post. A still at the exact boundary comes from video frames.

curl -X POST https://api.sume.com/v1/video-trim \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: insert-before-001" \
  -d '{
    "video_url": "https://media.sume.com/artifacts/artf_demo/source.mp4",
    "start": 0,
    "end": 12
  }'

Where does the manual route fall short?

Three places. The generated scene is billed at Sume's rate, not fal's: list times 1.25 per output second. Nothing in the docs promises the new scene matches the source's lighting or sound, so a hard cut can show. And the audio spine is yours to build; if the original has sound under the insert, you need to supply the replacement or silence yourself.

For a plain cut-in of a clip you already have, no generation is needed, and how to insert a video into another video covers that case with a single render.