MiniMax H3: multi-speaker voice/accent conditioning can leak between subjects in FL2VA and Ref2VA

### Custom Node Testing

- [x] I have tried disabling custom nodes and the issue persists (see [how to disable custom nodes](https://docs.comfy.org/troubleshooting/custom-node-issues#step-1%3A-test-with-all-custom-nodes-disabled) if you need help)

### Expected Behavior

In a MiniMax H3 multi-character dialogue scene, voice characteristics specified for one subject should remain bound to that subject.

For example, if Subject 1 is instructed to speak with a British voice and Subject 2 is instructed to speak with an American voice, Subject 1 should retain the British voice while Subject 2 retains the American voice.

Voice conditioning for one speaker should not alter the voice identity, accent, or vocal characteristics of another speaker in the same generated clip.

### Actual Behavior

In some two-character scenes, voice conditioning appears to affect the generated audio globally rather than remaining bound to the specified subject.

For example:

When Subject 2 is explicitly instructed to speak in an American voice, Subject 2 correctly becomes American, but Subject 1 also changes to an American voice.
When Subject 1 is then explicitly instructed to speak in a British voice resembling Keira Knightley as Elizabeth Swann from Pirates of the Caribbean, Subject 1 correctly changes to the requested British voice, but Subject 2 also changes to a different British voice.

The visual speaker assignment remains correct: the intended character speaks the intended line, lip synchronization and facial acting correspond to the correct speaker, and the characters remain visually distinct.

The apparent problem is specifically that voice/accent conditioning is not always remaining subject-local and can propagate to the other speaker.

I have reproduced the same behavior with both Ref2VA and FL2VA. Other character pairings can successfully maintain separate voices, so the issue does not occur universally.

### Steps to Reproduce

1. Load the MiniMax H3 model using the current ComfyUI H3 workflow.
2. Use FL2VA with a single first-frame image containing two clearly distinct female characters. No last frame is required.
3. In the prompt, identify the characters separately as and .
4. Assign different voice characteristics to the two subjects. For example:
◦ : young feminine British voice resembling Keira Knightley as Elizabeth Swann from Pirates of the Caribbean.
◦ : bright feminine American voice.
5. Give each subject one dialogue line in the same clip, for example an A → B exchange.
6. Generate the video.
7. Observe that the requested accent/voice characteristic may propagate between speakers. In my reproduction, specifying the American voice for Subject 2 caused both subjects to speak with American voices.
8. Modify the prompt so Subject 1's dialogue line explicitly states that it is spoken in the Elizabeth Swann / British voice.
9. Generate again.
10. Observe that Subject 1 now correctly uses the British voice, but Subject 2 also changes to a British voice rather than retaining the assigned American voice.
The issue has also been reproduced with Ref2VA and persists when using explicit dialogue tags, pure / labels, and only a single A → B speaker transition.

### Debug Logs

```powershell
None
```

### Other

_No response_

Start by reproducing the behavior with the current ComfyUI MiniMax H3 workflow using both FL2VA and Ref2VA, following the two-subject dialogue steps in the issue. Compare whether voice conditioning remains subject-local across repeated generations and with the listed tag formats; there are no files, tests, or debug logs named, so done would require identifying a reproducible cause and verifying that the two speakers retain their assigned voices.

Written by the indexing model from the issue text.