Katechon ArcWhat we saw
- In The Wolf Mother, Seedance sometimes invented a voice-over on near-empty shots such as a black frame. Replacing those clips with ambience-only takes kept the narrator the only voice in the film.
- In Dragon Rising, Seedance added a radio line nobody wrote in the opening. We dropped it from the captions, but it stays in the audio.
Why it happens
A model that generates picture and sound together has learned that video usually has people talking. When a shot gives it little to show (a black frame, an empty corridor, a slow landscape) and says nothing about sound, speech is a likely guess. A character who appears to be holding a radio, a phone or a microphone invites the same guess.
How to prevent it
- Say what the shot sounds like: wind and footsteps only, no dialogue, no narration. Positive descriptions of the sound work better than a bare ban.
- Give empty shots something to show, such as weather, movement or light changing, so the model has a subject other than a speaker.
- Keep props that imply speech (radios, phones, microphones) out of shots that should be silent.
- Put narration on its own track in the edit rather than asking the video model to leave room for it.
- Listen to every clip before captioning. A stray line that survives the edit should be either cut or captioned, never left unexplained.
When it does happen
Regenerate that one shot with an explicit ambience-only sound description; this usually fixes it in one or two takes. If the picture is perfect and only the audio is wrong, mute the clip and lay ambience or music under it in the edit. Either way, budget for it: these takes count toward the re-renders that cost us 15% of video spend across our first five series.
Where Katechon Arc fits
In Arc each shot's sound is written into the script, so you can mark shots as ambience only and regenerate a single clip without touching the rest of the episode.
Arc cannot strip an unwanted voice from a generated clip's audio; muting or replacing the sound happens in your editor.