Reproduce “MoST: Motion Style Transformer between Diverse Action Contents” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Boeun Kim, Jungho Kim, Hyung Jin Chang, Jin Young Choi

MoST, a transformer-based framework, advances motion style transfer by effectively disentangling style from diverse action content, producing high-quality, plausible stylized motions without heuristic post-processing. The framework achieves superior performance, demonstrating Content Consistency (CC) of 8.5 and Style Consistency (SC++) of 63.0 on the Xia dataset.

Key insights

  • A Siamese encoder coupled with a dedicated Style Disentanglement Loss (L_D) effectively separates style from content, enabling robust transfer even between diverse action types.
  • Part-attentive modulation within the PSM allows for nuanced style application, preserving the content motion's action while accurately reflecting the desired style across relevant body parts.
  • Directly integrating physics-based regularization, including novel foot contact loss, into the training process results in intrinsically plausible global translation and artifact-free motions without requiring any post-hoc adjustments.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.