Reproduce “StableMotion: Training Motion Cleanup Models with Unpaired Corrupted Data” on OpenResearch

About this paper

Yuxuan Mu, Hung Yu Ling, Yi Shi, Ismael Baira Ojeda, Pengcheng Xi, Chang Shu, Fabio Zinno, Xue Bin Peng

Researchers developed StableMotion, a framework for training motion cleanup models directly on unpaired, corrupted motion data using a diffusion-based approach with quality indicator variables. This method achieved a 68% reduction in motion pops and an 81% reduction in frozen frames on a proprietary soccer mocap dataset, while also outperforming state-of-the-art cleanup models on controlled benchmarks.

Key insights

• The introduction of frame-level quality indicator variables allows a diffusion model to effectively learn from unpaired, mixed-quality motion data, enabling explicit control over generated motion quality.
• A unified diffusion framework can simultaneously discriminate motion quality and generate cleaned motion through a masked inpainting objective, bypassing the need for paired supervision.
• Adaptive cleanup and ensemble inference strategies are crucial for improving content preservation, consistency, and robustness when cleaning motion with stochastic diffusion models.

Summary by alphaXiv.