Dai-Wenxun/MotionLCM

What it solves

MotionLCM addresses the runtime inefficiency of existing text-conditioned human motion generation models. While previous spatial-temporal control methods were slow, MotionLCM enables the generation of controllable human motions from text and control signals in real-time.

How it works

The project implements a motion latent consistency model (MotionLCM) built upon a motion latent diffusion model. It achieves real-time performance by using one-step or few-step inference. To maintain controllability, it integrates a motion ControlNet within the latent space and uses explicit control signals (such as initial motions) in the vanilla motion space to provide supervision during training.

Who it’s for

This tool is designed for researchers and developers working in human motion synthesis, computer animation, and AI-driven character control.

Highlights

• Real-time Generation: Uses latent consistency models to reduce inference steps to one or a few steps.
• Controllable Synthesis: Supports text-to-motion and motion control via a dedicated ControlNet.
• Comprehensive Pipeline: Includes tools for motion reconstruction, text-to-motion generation, and SMPL mesh rendering.
• Flexible Control: Allows control based on specific joints (e.g., the pelvis) or multiple joints simultaneously.