Talking Head Synthesis with Facial Landmark Guidance via 3D Gaussian Splatting

In plain English

Audio tells a model when sounds happen but not where the face is, so audio-only talking-head systems tend to produce sloppy mouth motion, weak expressions and local artefacts. This paper adds predicted facial landmarks in two places: to select and enrich 3D Gaussians around expression-sensitive regions, and as a whole-face conditioning vector that refines the Gaussian attributes. The authors report improved visual quality, facial realism and lip sync in both self-driven and cross-driven settings. For practitioners it is another data point that cheap landmark supervision remains an effective fix for the mouth-region failures audio-driven pipelines still have.