RIFT — Recoverability-Gated Implicit Facial Dynamics

I am investigating how much facial motion can be predicted from speech, and whether predictable motion is also the motion that matters for rendered articulation. RIFT began as a generative hypothesis about separating recoverable facial dynamics from residual uncertainty.

An initial predictive pilot did not support the tested formulation. I narrowed the project to RIFT-M: a measurement study using IMTalker’s frozen representation and renderer to examine the relationship between audio-predictable motion and changes in the rendered mouth.

Calibration exposed limits in the precision of standard landmark-based measurements. Before drawing conclusions about recoverability, I need to establish whether the measurement can reliably distinguish the changes being tested.

My work includes hypothesis formulation, controlled predictive experiments, failure analysis, and measurement design. The current priority is qualifying the measurement setup for broader evaluation. Improved generation and transfer across identities remain open questions.