Controllable Exaggeration for Generative Motion Models via Training-Time Adaptation and Inference-Time Guidance
The study developed a framework to make motion generative models produce more expressive character movements by incorporating animation principles, specifically exaggeration. This is achieved through two methods: fine-tuning models during training on an exaggerated dataset and using a mathematical formulation based on dynamic movement primitives during inference to guide the generation toward exaggerated motions.