Phantra AI: Sound Plan
Plans scene-specific diegetic sound, identifies models that render audio, and explains how to add music beds and recorded voice-overs.
Sound plan
A scene's sound is written into the prompt as DIEGETIC sound: what you would hear standing in that scene. Music, narration and captions are not in the prompt. Nothing in this plugin generates a voice or a song. Ambience can be added to a finished clip with an effect, quoted first like a render (effects).
Where each sound comes from
| sound | where it is made |
|---|---|
| what the scene itself makes (steps, a kettle, rain) | the prompt's sound line, rendered with the picture by a model that does sound |
| music | a bed under the clips: the studio's editor, or assemble_clips with music = an audio asset:<id> |
| a voice, narration, a spoken line | recorded by the person and laid on in the studio's editor |
| ambience or sound effects on a finished clip | an effect, quoted in credits first (effects) |
upload_link takes pictures and clips, never audio: a music file goes in
through the studio's editor, and library with kind audio lists it as an
asset:<id>.
The sound line
The scene ends, before the avoid list, with one line:
No background music. Only diegetic sound: <two to five sounds, in the
order they happen, each a thing on screen or just off it>.
Write sounds the picture explains: a footstep needs a foot, an engine a car. Say the room -- "close and dead", "a tiled echo", "wind over the mic". Never write a song, a score, a "cinematic swell" or a spoken line: the model either ignores it or invents a voice.
Per shot
The studio renders each timed window as its own clip. The closing sound line reaches every shot as the scene's memory, which is right when the sounds run through the whole scene. When the shots differ -- a kitchen, then a street -- give each window its own sentence opening "Sound:" after its action, and give every window one: a "Sound:" sentence on only the last shot is read as the scene's closing line, not as that shot's. The closing line then names what runs under all of them.
(0-4s) Mara lifts the lid of the pot and leans back from the steam. Sound:
the lid's clink, the boil, the hood's hum.
(4-10s) Hard cut to the doorstep: she steps out, pot in both hands, into the
rain. Sound: rain on the step, the door behind her, a car passing. End
abruptly.
No background music. Only diegetic sound: the pot, the rain, her breath.
Avoid: ...
Save with write_scene. A change to a saved scene goes in with its scene_id
and the whole new prompt: show the new version and get a yes first.
Which models render sound
video_models lists the models with the lengths, frames and credits each
takes. recommend_model with "with sound" in what ("a product clip with
sound") ranks the ones whose catalogue line says they render sound with the
picture, one reason each. A model not described that way is not promised to
have sound: plan for its sound to be laid on in the editor.
quote with model prices the scene on the one chosen and spends nothing:
keyframes (the stills to draw, each in credits), clip (one render per
timed shot, each with its credits and a token), balance and
credits_needed. Repeat the quote's own numbers -- each for a still, the
credits of each shot's render -- never a guess, and WAIT for a yes
in this chat; never say it for the person. Only then approve with the
scene_id, what ("keyframes" draws the stills, "clip" renders every shot
without one), the quote's tokens and the SAME provider, model, frame
and duration; poll job with wait, then scene shows the keyframe, the
clips and the scene's studio link. Tokens live one hour and work once; a
different renderer is a new quote and a new yes. Refused for
insufficient_credits? Stop, and say credits are added in the studio.
Music and voice, after the render
- Music:
assemble_clipsover the renderedgen:<id>s withmusic= the audioasset:<id>fromlibrary. The bed is trimmed to the picture and lowered under the clips' own sound. Free; polljobwithwait. The result is a newgen:<id>and an editable cut in the studio's editor, where the person balances it. For one clip alone, the editor. - The voice: the person records it (a phone in a quiet room is enough) and lays it on in the editor. Write the words with them if asked, as a script they read, never as a line in the prompt.
- Captions for those words: the editor's caption track, not the prompt.
Do not
- Never put music, a score, lyrics or a spoken line in the prompt.
- Never say a voice or a track was generated here; nothing is.
- Never pass audio through
upload_linkoradd_element. - Never promise a clip has sound unless the model's catalogue line says so.
- Never
approvewithout the price shown and a yes here, and never twice.
- Downloads
- Security audit
- Pass
- Last updated
- 2w ago
- Current version
- v1.0.0
- License
- MIT-0
