Developer seeks inference-time expression control for a frozen talking-face model

A builder of a photorealistic talking-face product is looking for ways to add expressiveness to a frozen diffusion adapter that synchronizes lips well but produces flat facial expressions. Retraining is ruled out, so the focus is on inference-time controllers. The post asks how others schedule blinks, brow movement, and head motion from audio prosody and phoneme timing, and how to keep slow expression signals from interfering with fast lip articulation. It also raises which failure modes, such as repetitive motion, lip interference, or flicker, appear first, and how to detect them in rendered clips. The author wants controller designs and evaluation methods rather than vendor recommendations.