MotionLCM presents a novel approach to real-time controllable motion generation, addressing the runtime inefficiency inherent in existing methods for text-conditioned motion synthesis. By building upon the motion latent diffusion model (MLD), MotionLCM introduces the motion ControlNet within the latent space to facilitate explicit control signals, enhancing the controllability of motion generation.

The model employs one-step (or few-step) inference techniques to significantly improve runtime efficiency, enabling the generation of human motions with text and control signals in real-time. Leveraging consistency distillation, MotionLCM accelerates motion generation without compromising quality, making it suitable for various real-time applications requiring precise motion synthesis.

To ensure effective controllability, MotionLCM incorporates a motion ControlNet that manipulates motion generation in the latent space, allowing direct influence of control signals on the generation process. Through optimization of parameters in the motion ControlNet and Trajectory Encoder, MotionLCM achieves superior performance in controllable motion generation, surpassing previous methods in text-motion alignment and motion control.