A visual reference number and a speaker number answer different questions. One identifies a referenced subject; the other identifies a vocal source. They do not need matching digits. If a prompt assumes that the third picture must use the third voice, an apparently correct cast can still deliver the wrong lines.
Make a role-to-voice crosswalk
The MiniMax H3 full-reference writing guide distinguishes referenced subjects from stable speaker IDs. This is an H3 rewrite convention. It is not the same thing as an attachment label such as @image1 in a website, and copying internal labels into an unrelated model does not establish support.
Before drafting shots, create a plain-language crosswalk. Include each person's visual source, story role, first spoken line and intended voice characteristics. Include a subject with no dialogue too: silence is useful ownership information. A room or product may also be a referenced subject without ever becoming a speaker.
Consider an invented scene with a room reference first, a courier second and a shopkeeper third. The shopkeeper asks the opening question, the courier answers, and the shopkeeper responds again. Subject ordering follows the reference definitions; vocal ordering follows actual speaking sources. Do not renumber the shopkeeper every time a new camera view appears.
Audit every line through both identities
| Role | Visual evidence | Vocal ownership |
|---|---|---|
| Shopkeeper | Internal <Subject 3> | (S1): opening question and final response |
| Courier | Internal <Subject 2> | (S2): answer only |
| Shop interior | Internal <Subject 1> | No human voice assigned |
Check pronouns after cuts. A sentence such as “he answers” becomes ambiguous if both men are present or the shot has switched. Restore the role name in the working brief, then use the actual H3 convention consistently in a supported rewrite pipeline. The full-reference structure contains subject definitions, summary, retention analysis, detailed description, soundscape and music; do not submit an isolated mapping as though it were a complete request.
Keep dialogue wording separate from descriptions of tone or movement. A stage direction accidentally placed inside spoken text can be voiced as dialogue. A narrator introduced later needs a stable independent vocal identity rather than borrowing the nearest visible character's voice.
Review voice ownership before sound quality
Read only the spoken lines in order and ask who must produce each one. Then inspect the corresponding visible person or off-screen designation. This paper review can catch a contradictory assignment without spending on a generation, although it cannot prove that the eventual output will obey it.
After generation, listen to each line and inspect the intended speaker's mouth while checking that listeners remain listeners. A pleasant voice is still wrong if it belongs to the wrong role. Keep the crosswalk with an Imagild recreation brief, using the site's actual numbered attachments for user-facing instructions. Treat any internal provider rewrite as a separate representation that must preserve that same ownership.