Decide whether the music exists inside the scene or only in the audience's soundtrack. “Soft piano while she waits” leaves that ownership unclear: it could mean an unseen pianist, a radio or an editorial score. If the character should neither perform nor react to the music, describe it as audience-only accompaniment and keep it separate from physical scene events.
Ask who can hear each sound
The MiniMax H3 base guide distinguishes sound in the story world from non-diegetic music. Its rewrite structure gives audience-only score its own field, while music produced by a radio, instrument or person belongs to the audiovisual scene description. This is H3-specific guidance, not a guarantee that any chosen model exposes separate audio tracks.
Create a small sound-source table. State who produces a sound, who hears it and whether any visual response is intended. Ordinary room ambience can exist without a visible appliance in every shot; the point is to avoid assigning a physical performance that the story never requested.
Compare two deliberately different versions
In an invented scene, a ceramic artist waits for a kiln to cool. Version A uses a sparse instrumental score for the audience. The artist watches the temperature display and does not play an instrument. Version B has an actual radio on the shelf; the artist turns its volume down before speaking. They may share a musical mood, but the required actions differ.
| Sound intention | Necessary scene evidence |
|---|---|
| Audience-only score | No performer or listening reaction implied |
| Shelf radio | Consistent source and any requested interaction |
| Character singing | Approved lyrics and visible vocal performance |
For an H3 base rewrite, preserve the full three-field structure rather than dropping a lone music instruction into an unrelated request format. Describe concrete instrumentation and progression when known. Avoid inventing a piano in the room simply because the score uses piano sounds.
Inspect ownership before changing the mix
Watch the result muted first. Has a musical instrument appeared, has the character begun singing, or is a gesture now synchronized as an unrequested performance? Those are visual consequences of source ownership. Lowering the soundtrack volume cannot remove them.
Then listen with the picture. Check whether scene sounds still exist under the score and whether an intended radio sounds like a plausible part of that room. Do not infer a specific speaker position or separate stem delivery unless the tool actually supplies it. The comparison described here is a planning exercise, not a published model test.
Put the intended sound ownership in an Imagild film brief, especially when preserving a reference soundtrack. Accept when the requested music role and the visible action agree. If the actual output turns an audience score into character performance, reject that interpretation even if the melody is attractive. Repair the ownership brief or use an approved silent picture with externally added music where the workflow permits.