No Distillation Needed: Single-Pass Real-Time Talking Heads via Acausal Noise Shaping

In plain English

The authors argue that audio-driven facial animation does not need diffusion models at all. Their FaceGAN generates expression and head pose one frame at a time in a single pass, with the noise input shaped acausally to avoid a known failure mode of causal generators, and they report quality matching or beating diffusion-based state of the art. For practitioners the interesting claim is the bounded-attention design, which they say lets the model run indefinitely without drifting — the usual reason long streaming sessions degrade.