Generative Motion In-Betweening by Diffusion Over Continuous Implicit Representations

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Generative models show progress in motion in-betweening for realistic transitions.
• Existing methods struggle with keyframe preservation and motion continuity.

Purpose of the Study:

• To propose a novel pipeline and sampling strategy for latent diffusion models (LDM) using motion implicit neural representations (INR).
• To improve motion generation quality with sparse keyframe data.

Main Methods:

• Developed a mapping between INR and sparse spatial/temporal information in latent diffusion.
• Sampled INR parameters from sparse keyframe data to reconstruct motions.
• Utilized a novel sampling optimization strategy.

Main Results:

• Achieved superior performance in motion generation with few keyframes.
• Significantly improved motion generation quality.
• Ensured both keyframe accuracy and diversity of in-between motions.

Conclusions:

• The proposed LDM-based pipeline with INR effectively reconstructs plausible and smooth motions from sparse data.
• This method addresses limitations in keyframe preservation and continuity for motion in-betweening.
• Demonstrated enhanced motion generation quality and diversity.