Controllable human motion generation

Abstract

Controllability is crucial in text-to-motion generation because it allows users to specify precise motion details, such as the trajectory of a limb or the orientation of a body part. This level of control is vital for applications in fields like gaming, where unique character movements enhance user experience, or in animation, where users can manipulate motions to suit narrative or aesthetic needs. However, existing models often face limitations in balancing semantic fidelity, control, and efficiency. For instance, while some methods offer fine-grained control over joint movements or body parts, they may require extensive computation, reducing their practicality for real-time applications. This project examines the impact of various samplers on improving inference speed for OmniControl, a controllable motion generation approach. Additionally, a graphical user interface has been developed as a functional prototype.