Generating Multimodal Human Dynamics from Transformation-regularized Representations