Avatar Video Scene Durations: Fields Must Agree and the 2.8 Rule

In a Sume Avatar 1.0 multi-scene request, a text scene without an explicit duration is estimated at ceil(words / 2.8) seconds, with a minimum of 1 second. If you do set a duration in more than one field, they must agree within 0.001 seconds or the request is refused. The total must still land at 4 to 60 seconds.

This is from Sume's Generate avatar video docs and the implementation, read on 2026-10-05.

How is the length of a scene worked out?

Words are split on whitespace, divided by 2.8 (the words-per-second constant), and rounded up. That means a short line costs a full second: the estimate is a planning number, not a measure of how long speech takes to say.

Billing for Avatar Video is per second of the video at the quality tier's rate, so a longer estimate costs more. At the standard rate of $0.184 per second, a 5-second scene is $0.92 and a 6-second scene is $1.10.

Which duration fields must agree?

A scene can carry its length in several places: voice.duration, duration_seconds, duration, or a timing pair of start_seconds and end_seconds. If you provide more than one, the values must match within 0.001. A timing pair needs both ends, and end_seconds must be greater than start_seconds. No scene may exceed 60 seconds.

Silence scenes are different: voice.type: "silence" requires a duration and permits no script.

{
  "id": "beat",
  "voice": { "type": "silence", "duration": 2 }
}

What should you do?

Set one duration field per scene and leave the rest off. For spoken scenes, let the 2.8 rule decide unless you have a reason to fix the length, and keep a spare second in the 4 to 60 window so a late edit does not push you over.