Reality Check: How Avatar and Face Representation Affect the Perceptual Evaluation of Synthesized Gestures

In plain English

Three experiments with 131, 180 and 164 participants show that how you render an avatar changes the verdict viewers reach about the gestures it performs. A moving face inflated speech-gesture match ratings and a masked face was the most distracting, so the authors recommend a blurred face as the least biased baseline for gesture studies. The avatar body mattered just as much: a Gaussian-splat avatar separated motion-captured from synthetic motion most sharply (d = 2.65) while a stick figure barely did (d = 0.88). The practical implication is that gesture-generation results evaluated on cheap or abstract avatars will understate the gap between synthetic motion and real performance.