Avatar script over 60 seconds: split it into jobs and join the clips

Sume rejects an avatar script whose estimated duration is outside 4 to 60 seconds, and the docs say what to do: shorten it or split it into multiple jobs. For a three-minute talking-head explainer that means about four to six jobs, one per chapter, each with its own Idempotency-Key, rendered with the same avatar_handle, quality and aspect_ratio so the parts match.

Sume gives you the parts; stitching them into one file is your editor's job. Sume's Avatar route does not concatenate jobs.

Plan the split yourself

Split at sentence ends, not mid-sentence, and keep each part well under the limit. The words-per-second figure below is a planning assumption of this example, not a Sume number; the real check is Sume's own duration estimate, so a part can still be rejected.

import re

WPS = 2.2        # planning assumption, not a Sume figure
MAX_SECONDS = 50 # leave headroom under the 60 s limit

def split_script(text):
    parts, cur = [], []
    for s in re.split(r"(?<=[.!?])\s+", text.strip()):
        trial = " ".join(cur + [s])
        if cur and len(trial.split()) / WPS > MAX_SECONDS:
            parts.append(" ".join(cur))
            cur = [s]
        else:
            cur.append(s)
    if cur:
        parts.append(" ".join(cur))
    return parts

parts = split_script("One. " * 300)
print(len(parts), [len(p.split()) for p in parts])

Keep the parts consistent

• Use one avatar_handle, one quality, one aspect_ratio for every part.
• Key each job by part: explainer:v1:part-01, part-02, and so on. A retry of part 3 then cannot create a second part 3.
• Submit all parts, then poll each GET /v1/jobs/{id}/status. The results are separate media.sume.com video artifacts.
• If a part is under 4 seconds it is rejected, so merge a one-line ending into the part before it.

A table of the limits

What Sume does not do

There is no "long video" mode for Avatar 1.0 in the docs, and no join endpoint for the parts. Join the finished parts in your own editor, or with a Timeline render if your workflow already uses one.