Real-time controllable motion transition for characters

Real-time in-between motion generation is universally required in games and highly desirable existing animation pipelines. Its core challenge lies the need to satisfy three critical conditions simultaneously: quality, controllability speed , which renders any methods that offline computation (or post-processing) or cannot incorporate (often unpredictable) user control undesirable. To this end, we propose a new real-time transition method address aforementioned challenges. Our approach consists of two key components: manifold conditional transitioning. The former learns important low-level features their dynamics; while latter synthesizes transitions conditioned on target frame desired duration. We first learn explicitly models intrinsic stochasticity human motions via multi-modal mapping mechanism. Then, during generation, design model essentially sampling strategy sample from learned manifold, based aimed validate our different datasets tasks where no post-processing allowed. Through exhaustive evaluation comparison, show able generate high-quality measured under multiple metrics. also robust various frames (with extreme cases).