Efficient stochastic human motion prediction via consistency model

Diffusion models have recently proven successful in stochastic human motion prediction. Despite their excellent generative performance, they are difficult to predict in real-time because the multi-step sampling mechanism involves tens or hundreds of iterations of repetitive function evaluation. To address this issue, we introduce the Motion Consistency Model (MotionCM) to alleviate the computational and time consumption of the iterative inference process. It applies a diffusion pipeline to the low-frequency domain processed by discrete cosine transformation(DCT), alleviating the computational burden for each function evaluation. The predictive efficiency of the motion diffusion model in stochastic human motion is further improved by employing a one-step (or several-step) inference via maintaining consistency of the outputs over the trajectory of PF-ODE. Extensive experiments on two benchmark datasets show that the proposed model achieves state-of-the-art performance at less than 10% of the time cost.