Maestro v3.2.5: Face Refinement Causes Incorrect Lip-Sync in Two-Person Dialogue
I discovered an issue with Face Refinement while generating a two-character conversation using H3 Omni Pruned.
In the original generated video, everything works correctly. When Character A speaks, Character B listens silently with their mouth closed, and vice versa.
However, when Face Refinement processes that same video, both characters begin moving their mouths in sync with the dialogue, even though only one character is speaking.
The interesting part is that Maestro saves both versions separately, allowing a direct comparison. The verbal hallucinations at the end of the dialogue, is a whole different issue.
What I observed:
- Original video: Correct speaker-specific lip movements.
- Face-refined video: Both characters mouth the same dialogue.
- Audio and dialogue remain the same.
- The problem appears during Face Refinement, not the initial H3 generation.
Setup: Maestro v3.2.5, H3 Omni Pruned, 540p, RTX 4070 Super 12GB.
I spent quite a while experimenting with speaker assignments, audio references, and prompt instructions before comparing the original and refined outputs.
Workaround: Disable Face Refinement for two-character dialogue scenes, or use the original generated video.
Hopefully this helps identify the issue. I've provided both video versions for comparison.
Maestro v3.2.5: Face Refinement Causes Incorrect Lip-Sync in Two-Person Dialogue
I discovered an issue with Face Refinement while generating a two-character conversation using H3 Omni Pruned.
In the original generated video, everything works correctly. When Character A speaks, Character B listens silently with their mouth closed, and vice versa.
However, when Face Refinement processes that same video, both characters begin moving their mouths in sync with the dialogue, even though only one character is speaking.
The interesting part is that Maestro saves both versions separately, allowing a direct comparison. The verbal hallucinations at the end of the dialogue, is a whole different issue.
What I observed:
Setup: Maestro v3.2.5, H3 Omni Pruned, 540p, RTX 4070 Super 12GB.
I spent quite a while experimenting with speaker assignments, audio references, and prompt instructions before comparing the original and refined outputs.
Workaround: Disable Face Refinement for two-character dialogue scenes, or use the original generated video.
Hopefully this helps identify the issue. I've provided both video versions for comparison.