Reproduce “DiffusionPhase: Motion Diffusion in Frequency Domain” 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
Weilin Wan, Yiming Huang, Shutong Wu, Taku Komura, Wenping Wang, Dinesh Jayaraman, Lingjie Liu
DiffusionPhase synthesizes realistic human motion from text prompts by operating in the frequency domain, addressing limitations in generating long, arbitrary-length sequences and achieving smooth transitions. The method maintains consistent quality for motions up to 5x the training length, outperforming existing time-domain diffusion models.
Key insights
- Representing motion in the frequency domain through a learned periodic phase space allows for compact encoding and arbitrary-length generation.
- The explicit inclusion of high-frequency components in the periodic representation enables capturing fine-grained motion details often missed by prior phase-based methods.
- Conditioning the diffusion model on both text prompts and starting poses facilitates smooth, natural transitions between different motion segments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.