Progressive Disentangled Representation Learning for Fine-Grained Controllable Talking Head Synthesis

Figure 1. Our method takes an appearance reference as input and generates its talking head with disentangled control over lip motion, head pose, eye gaze&blink, and emotional expression, where the driving signal of lip motion comes from speech audio, and all other motions are controlled by different videos. As shown, it well disentangles all motion factors and achieves precise control over individual motion.