AI video first frames jump in size? Trim two frames with source_in

If a 1080p image-to-video clip from MiniMax H3 Max jumps in crop or proportion right after its first two frames, start its Timeline slot at the third frame. For 24 fps clips that is source_in: 0.08333333333333333, with start and duration unchanged. Sume's model notes call this a measured workaround, not a rule for every clip.

The symptom

The H3 Max model card in the Video Router (apps/api/src/videos/sume-router/model-cards/minimax-h3-max.md) documents it: some 1080p image-to-video outputs have a crop or proportion jump after their first two decoded frames (issue #7429). The change is baked into the pixels. Setting a fixed frame size on the player does not hide it, and a Timeline fit: "cover" cannot remove it.

Confirm before you trim

Extract the first six frames of the raw clip and compare them by eye, or run the clip through Sume's frame tools. If frames 0 and 1 differ in framing from frame 2 onwards, you have the transient. Different source resolutions alone are not evidence of the bug, and talk clips, other models and clips with stable openings should not be trimmed by association.

The Timeline fix

Timeline 1.0 slots take a source_in, the in-point into the file. Two frames at 24 fps is 2/24 = 0.0833 seconds. Keep the slot's start and duration as they were, and check that the remaining source still covers the slot plus any transition you put on the following slot. Then render and look at the first displayed frames again.

{
  "audio": { "url": "https://media.sume.com/artifacts/artf_demo/voice.wav", "duration_seconds": 10 },
  "video": [
    {
      "source_url": "https://media.sume.com/artifacts/artf_demo/h3max.mp4",
      "start": 0,
      "duration": 10,
      "source_in": 0.08333333333333333
    }
  ]
}

Budget and limits

A Timeline render costs $0.10 per output minute, rounded up, and runs on worker ffmpeg with no provider inference. The plan endpoint compiles the document without creating a job, so you can check the timing first (see our plan note). Every source_url must be a media.sume.com artifact or asset in your workspace, so import the clip first.

If the clip is 10 seconds and you trim 0.083 s, make sure the slot duration does not exceed what is left. A 10.00 s source minus the trim covers 9.917 s, so shorten the slot or let the slot duration be 9.9. Timeline pads short sources and reports a soft warning rather than failing, but a visible freeze at the end is worse than a clean cut.

Other models: Sume's notes make no such claim for Wan 3.0, Seedance 2.5 or Omni, so do not apply the trim there.

Why it matters in a cut

A two-frame jump is 83 ms at 24 fps. In a hard cut it reads as a flash on the first beat of a shot, and with a crossfade it can show as a brief zoom as the fade starts. Because Timeline fades only apply on slots after the first, the first slot of a sequence shows the problem unmasked.

Keep the original, untrimmed file. The trim lives in the Timeline document, so you can remove it later if the provider fixes the output.