ViMoGen: A Novel Motion Generator for Virtual Standard Patient

Virtual Standard Patient (VSP) is an indispensable tool in medical education, offering crucial experiential learning opportunities. However, current VSP systems struggle to accurately represent patient behaviors and symptoms necessary for real-world diagnostic and assessment tasks. To address this challenge, we introduce ViMoGen, a novel motion generator for VSP driven by a controllable generation pipeline. Specifically, we introduce a conditional control mechanism for our diffusion-based generator. It is guided by dual inputs: instructional text prompts that simulate a physician’s commands, and more critically, expert-defined spatial constraints on specific body joints related to symptoms. This primary contribution allows for the direct encoding of physical limitations, ensuring the generated motions are medically grounded. At the same time, to further enhance the fidelity of the output, we introduce a task-specific loss guidance mechanism, this module refines the initially generated motion by leveraging targeted distance and absolute position losses. This optimization step ensures greater physical plausibility and precision in the final animation. Our experiments demonstrate that by synergistically combining conditional joint control and loss-guided refinement, ViMoGen produces realistic, fine-grained, and medically consistent body motions, making it highly suitable for disease research and medical training scenarios where the interplay of verbal and nonverbal cues is paramount.