Key points
- DeepPhase learns a compact phase manifold from motion capture data using a periodic autoencoder, enabling smooth, controllable character locomotion without hand-crafted phase labels.
- The method received the SIGGRAPH 2022 Best Paper award and is the current reference for neural phase-based character control in interactive and real-time animation contexts.
- DeepPhase is the current state-of-the-art baseline for neural phase-based character locomotion, used as a reference in subsequent character-control and motion-generation research.
Summary
Starke, Mason, and Komura (2022) introduce a periodic autoencoder that discovers the phase structure of motion data automatically, learning a continuous manifold that supports smooth transitions between locomotion states without requiring manually annotated phase signals. The method received the SIGGRAPH 2022 Best Paper award and is animation-native, having been designed and evaluated entirely within the character animation domain. It is the current reference for neural phase-based character control and has demonstrably shaped the direction of interactive character locomotion research.
Related items
- A deep learning framework for character motion synthesis and editing (SIGGRAPH 2016 / SIGGRAPH 2026 Test-of-Time award)
- Human Motion Diffusion Model (MDM)
- Iterative Motion Editing with Natural Language
Source
Source: ACM Transactions on Graphics (SIGGRAPH 2022) ↗ (Research)
Cite this item
Starke, S., Mason, I. & Komura, T. (2022). ‘DeepPhase: Periodic Autoencoders for Learning Motion Phase Manifolds’, ACM Transactions on Graphics (SIGGRAPH 2022). Available at: https://dl.acm.org/doi/10.1145/3528223.3530178
Your reference manager can also read this page directly: with the Zotero (or Mendeley) browser connector installed, save it straight to your library. Whole-collection exports: RIS, BibTeX, CSL-JSON.