Feature Request: User-Controlled Audio Modes for Video Generation

The new Wan 3.0 and MiniMax H3 related models essentially add powerful native audio capabilities that will bring video generation to the next level. At this early stage, though, audio generation can be rather unpredictable and counter-productive: clips may contain unsolicited ambient sounds, music, vocalizations/gibberish, or unprompted dialogue. More subtly, speech-like sounds (e.g. in a vocal background track) may trigger lip synchronization although no speech (e.g. a monologue or dialogue) was intended.

It would therefore be very helpful to give users explicit control over what kind of audio generation is permitted for a clip. One possible structure could be:

1. OFF / Silent — No generated audio: no ambient or action sounds, music, vocals, or dialogue. Importantly, this should also suppress speech-related lip synchronization. Natural facial and mouth movements (e.g. to express surprise or as part of other expressive gestures) should remain possible, but characters should not appear to articulate speech when none was prompted by the user.

2. NATURAL / Scene-Consistent — Generate only sounds naturally implied by the depicted scene and visible events: footsteps on the actual surface shown, a closing door, environmental ambience, etc. No unsolicited dialogue, vocals, or background music.

3. DIRECTED — Audio becomes an explicit and intended part of the user’s generation instructions, e.g. short specified dialogue (“Character A says: ...”), “let the character comment on the scenery/action in its own words,” or “add a calm, jazz-flavored instrumental background track.” Scene-implied natural sounds are not automatically included in this mode; users who want both would select the fourth option.

4. NATURAL + DIRECTED — Combine scene-implied natural sound with explicitly user-instructed audio components specified in the prompt. Here, the generation instructions would typically specify how both audio layers (scene-implied and user-instructed) should interact, blend, and persist over the clip’s timeline and, where applicable, across clip continuations. This is key to avoiding a disconnected audio mosaic.

(At the more advanced end, a production workflow could use externally supplied audio – e.g. a pre-generated voice track and/or soundtrack – as authoritative inputs, with the video synchronized accordingly through lip movements, facial expression, body language, etc. This goes substantially beyond the basic DIRECTED mode meant here, but helps delineate its scope.)

From a UI perspective, one possible implementation could simply be an Audio Mode control with four mutually exclusive options:

OFF | NATURAL | DIRECTED | NATURAL + DIRECTED

Depending on what the respective native model/API can reliably support, only a subset might initially be available. Even a robust OFF option would already provide substantial value at the start.

Feature request in a nutshell:

Explicit user control over the permitted audio generation through four mutually exclusive modes: OFF | NATURAL | DIRECTED | NATURAL + DIRECTED — i.e. no audio | scene-implied natural sound only | explicitly user-instructed audio only | scene-implied + user-instructed audio (typically, with user-guided balancing of both components)

If you think this is worth exploring by the Dev team, please forward it. Happy to be more specific if that helps. Together with the recent ⟨⟨Reference-Guided Video Continuation request⟩⟩ (2026-09-17), this could further strengthen and unify user control over the new generation of video models, both for single clips and connected sequences of clips.

Best regards,
Scotti