TL;DR of the paper, generated automatically.

A paper on restoring the mouths and faces in AI-generated videos, with the speech audio steering the lips and teeth.

• What it does: Sharpens small facial regions and aligns lip movement with the audio.
• How: A video diffusion transformer mixes the source video with noise, guided by reference images and earlier frames.
• Claims: On talking-head benchmarks, the authors report the lowest LPIPS and DISTS and the best lip sync among the restoration baselines they tested.
• Also: It supports visual dubbing when the audio is replaced.