Reproduce “Back to Basics: Motion Representation Matters for Human Motion Generation Using Diffusion Model” on OpenResearch

About this paper

Yuduo Jin, Brandon Haworth

An empirical study clarifies the impact of motion representations and loss functions on human motion generation using diffusion models, finding that a simple Joint Position representation combined with v-loss and geometric regularization yields superior motion quality and diversity while substantially reducing training time.

Key insights

• The simpler Joint Position (JP) representation, particularly when combined with v-loss and geometric regularization, consistently outperforms complex rotation-based representations in terms of generative quality and diversity within the vMDM framework.
• Incorporating geometric losses (position, velocity, foot contact) alongside the v-loss function is crucial for achieving high-fidelity and stable motion generation, as relying solely on v-loss degrades quality.
• Choosing the Joint Position representation significantly reduces training time by avoiding the computational overhead of forward kinematics calculations required for rotation-based representations.

Summary by alphaXiv.