Motion Field

Motion fields represent how each point in an image should move to transform from one state to another, encoded as a 2D displacement map. In motion-based lip sync approaches, rather than directly generating pixels, the model predicts a motion field that warps the source face into the desired mouth position.

This approach preserves original face texture and identity better than direct pixel generation because it moves existing pixels rather than creating new ones from scratch. Models like face-vid2vid use dense motion fields combined with neural rendering to achieve high-fidelity lip sync with strong identity preservation.

Motion Field relates to several other concepts in the AI lip sync pipeline: Optical Flow , and Face-vid2vid .