Wenxun Dai Ling-Hao Chen Jingbo Wang Jinpeng Liu Bo Dai Yansong Tang

Abstract

This work introduces MotionLCM, extending controllable motion generation to areal-time level. Existing methods for spatial control in text-conditionedmotion generation suffer from significant runtime inefficiency. To address thisissue, we first propose the motion latent consistency model (MotionLCM) formotion generation, building upon the latent diffusion model (MLD). By employingone-step (or few-step) inference, we further improve the runtime efficiency ofthe motion latent diffusion model for motion generation. To ensure effectivecontrollability, we incorporate a motion ControlNet within the latent space ofMotionLCM and enable explicit control signals (e.g., pelvis trajectory) in thevanilla motion space to control the generation process directly, similar tocontrolling other latent-free diffusion models for motion generation. Byemploying these techniques, our approach can generate human motions with textand control signals in real-time. Experimental results demonstrate theremarkable generation and controlling capabilities of MotionLCM whilemaintaining real-time runtime efficiency.