Lip sync audio under 5 seconds: H3 Max rejects it, use Fabric

A 4.5-second line sent to Sume's MiniMax H3 Max lip sync is refused with invalid_request: its duration_seconds window is 5 to 14.8, and Sume never clamps a value into it. The fix is to send that clip to VEED Fabric 1.0, which takes the same body and accepts 1 to 300 seconds on its image-to-video route.

The same applies above the window: 15 seconds is refused too.

Why is the window 5 to 14.8 seconds?

It is the provider's. Sume's lip sync note says the upstream rejects audio under 5 seconds and silently clips anything past 14.8 seconds. A clamped reservation would then render a clip shorter than your audio and drop speech, so Sume checks duration_seconds itself and refuses an out-of-range value rather than changing it.

duration_seconds is required and is the basis for the reservation, not a trim instruction. Measure the real audio and send that number.

What does the call look like?

The request is the same for both models: a visual source (image_url or avatar_handle, never both), an audio_url hosted on Sume, duration_seconds, and an optional resolution. Switch the path, not the body: POST /v1/minimax/h3-max/lip-sync for H3 Max and POST /v1/veed/fabric-1.0 for Fabric.

curl -X POST https://api.sume.com/v1/veed/fabric-1.0 \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: short-line-004" \
  -d '{
    "image_url": "https://example.com/host-still.png",
    "audio_url": "https://media.sume.com/example/line-4-5s.mp3",
    "duration_seconds": 4.5
  }'

Which model for which length?

Use the audio length to choose before you submit. Do not pad a short line with silence or cut a long one just to enter the H3 Max window; Sume's guidance is never to re-cut or re-synthesize audio to fit a model.

If you are building an agent, let it measure the duration and route: a clip inside the window may use H3 Max, and everything else goes to Fabric.

What about the 400 and the bill?

An invalid_request is an immediate rejection: Sume only accepts a job when the request can be safely accepted, so nothing is queued. Fix the model or the duration and resubmit. Use a new Idempotency-Key when you change the body, because reusing a key for a different payload returns 409 idempotency_conflict.

If you retry after a timeout, keep the same key with the identical body, and poll the job you already have.

How do I measure the duration reliably?

Probe the file, not the script. A script's spoken length is an estimate; the audio file's real length is a number. Any audio probe gives you the figure to send as duration_seconds; Sume's own Video inspect does the same for a hosted clip.

Send the measured value to the decimal, such as 4.5, not a rounded-down one: duration_seconds is the basis for the price reserved at submit, so an understated number reserves too little and hides the real cost.