A 58-word six-sentence avatar script becomes a 10 s and a 12 s clip

A 58-word script made of six sentences (9, 9, 9, 10, 10 and 11 words) is planned by the Sume Avatar 1.0 chunker as two clips: the first holds the opening three sentences, 27 words, planned at 10 seconds, and the second holds the last three, 31 words, planned at 12 seconds. The total is 22 seconds. The cut falls after the third sentence because adding the fourth would take the first clip to 37 words, above the 33-word ceiling.

How the packing works

The numbers on this page come from running the avatar-workflows package in the Sume repo on constructed scripts, and from the Avatar videos docs for the 4 to 60 second window. The chunker is an internal planner: it decides how many clips your script becomes, and the public docs describe the outcome rather than the algorithm, so treat the exact constants as current behavior that can change.

Sentences are packed greedily in order. The planner keeps adding the next sentence to the current clip until the next one would push it past 33 words, as long as the current clip has at least 7 words. Then it starts a new clip. Nothing is moved backward, so the split is not balanced across clips: here it is 27 and 31, not 29 and 29.

Reading the durations

Clip 1: 27 divided by 2.8 is 9.6, rounded up to 10. Clip 2: 31 divided by 2.8 is 11.07, rounded up to 12. The two add to 22, which is inside the 4 to 60 second window of the talking video. Note that the 12 is also the maximum for a clip; one more word per sentence at the end would have forced a third clip rather than a longer second one.

Use it to edit a script

Because the packing is greedy, the end of each sentence is where you control the cut. A few habits follow.

• Aim for sentences that sum to about 25 to 33 words so each clip uses its time.
• If two short sentences fall on either side of a cut, they will be in different clips, which means a new generation and possible small differences in pose.
• Avoid ending on a very short sentence; a tail under 7 words is a special case in the planner.
• Compare the chunk count the job reports with this plan.

What Sume does not do

The chunker does not look at meaning. It will cut between a question and its answer if the word count says so. If you want a specific break, make it a scene boundary. The planner also assumes English-like word spacing; the sentence splitter does recognize full-width punctuation, but a language without spaces will not be counted the same way.