Tiny is not small enough: High-quality, low-resource facial animation models through hybrid knowledge distillation

Researchers have developed a breakthrough method for creating high-quality, speech-driven 3D facial animation models that are small enough to run in real-time on devices. Using hybrid knowledge distillation, they trained tiny 'student' models from a larger 'teacher' model, eliminating the need for massive datasets. The resulting models require only 3.4 MB of memory and 81 ms of future audio context while maintaining animation quality, enabling on-device inference for games and digital characters.