Training video policy change: re-render one 15 s module

If one sentence in a 45-second avatar training video changes, you re-render all 45 seconds, which costs $8.28 on Sume's standard tier. If the video is built as three 15-second modules joined afterwards, you re-render only the module that changed for $2.76, plus a timeline render at $0.10 per output minute and a $0.01 audio detach to join them again.

Prices are from Sume API pricing, the timeline rate from Timeline 1.0, and the avatar video rules from Generate avatar video, read 2026-10-03.

Why a single clip is expensive to maintain

A Sume avatar video is one avatar speaking one script or one multi-scene plan, and the request is one job. The current execution supports one resolved avatar per final video, with backgrounds that resolve to one shared scene. The docs describe no way to patch a sentence in a finished clip, so changing the words means a new render of the whole request.

For a video that is stable for years, that is fine. For compliance training that changes with each policy update, the repeat cost is the hidden number.

How to modularise

Split the video by topic, not by length. A module is one policy, one procedure or one warning, scripted to run about 15 seconds, so a change touches exactly one job. Keep the avatar handle, aspect ratio, quality and background prompt identical in every module so the scenes match when they are joined.

Join the finished modules with a timeline render, which is priced at $0.10 per output minute, rounded up to the minute. A 45-second joined video is one billable minute. A timeline render needs an audio spine, so detach the audio from the new module ($0.01 per job, see Audio detach) and pass it with the kept modules' audio as audio.parts[]. Re-joining after a module changes costs $0.11 in all, which is small next to the avatar saving.

• Name each module's key by topic and version, such as training-fire-exit-v3.
• Store the job id of each module so you can swap one in the timeline later.
• Keep a script file per module in version control, so the change history is the script history.

What this does not fix

Modules add joins, and a join can show where scenes meet: a slight difference in framing, lighting or pacing. Preview the first frame of each module, and keep the background prompt exactly the same. If the video must read as one continuous take, the single clip is the safer design, and you accept the re-render cost.

Also decide who approves a change. A re-rendered module is new generated media, so your review step should check the output, not only the script, before the joined video replaces the old one in your learning system.

The saving grows with every change. A year with four policy updates to one topic costs 4 x $8.28 = $33.12 in re-renders of a single 45-second video on standard, against 4 x ($2.76 + $0.10 + $0.01) = $11.48 with modules, ignoring the first build. That arithmetic uses list rates and assumes each change touches exactly one module, which is the design goal.