In the app’s walkthrough: Tools · Optional sound pass
Before you begin
Select a small set of unlocked takes. Check that their visual direction describes the action correctly before asking a model to write sound cues.
Work through it
Choose Tools → Add Sound Effects
This prepares the selected takes before video rendering. It adds a SOUND EFFECTS FOR THIS RENDER block to their saved prompts, including Silent + VO takes.
Choose Write sound cues with
Select a particular script model, the chapter script model, or No LLM — generic Foley direction. A model writes scene-specific cue lines; the generic option does not call a text model.
Preserve or rewrite existing cues
Leave Rewrite cues that takes already have off to keep existing audio directions. Turn it on only when you intend to replace them with newly written cues.
Read the saved prompt
Check the result on one take. Useful cue wording names sounds directly, for example “Faint breeze and distant machinery hum.” Remove irrelevant sounds before rendering.
Render and listen to the layers
Review the renderer’s effects against the original recording. For takes without lip sync, the vocal-removal pass uses Demucs locally to reduce invented voices while retaining music and effects. Its setup can require several GB; without it, those takes finish with your recording only.
Remove the effects instruction when needed
Tools → Remove Sound Effects removes the render-prompt block and requests a render without effects. It keeps saved cues for later reuse. Changing the prompt does not rewrite a video already rendered.
Try it in your chapter
Add cues to one take, inspect the prompt, and compare the rendered audio with the source recording. Listen for stray speech or damaged effects.
Check before moving on
The effects support the scene, the intended narration remains audible, and you know which layers you want in the export.
If something goes wrong
The task updates no takes
Check locks and whether the takes already have cues with rewriting disabled. Read the completion message rather than assuming the task is still running.
The video still contains invented speech
Check vocal-removal setup and listen to its output. Separation can leave artifacts; use the original recording alone if the remaining track is unsuitable.
Still stuck? Contact support with the exact message and the operation you were running.