Reproduce “TEDi: Temporally-Entangled Diffusion for Long-Term Motion Synthesis” 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
Zihan Zhang, Richard Liu, Kfir Aberman, Rana Hanocka
The TEDi framework extends Denoising Diffusion Probabilistic Models to generate arbitrarily long, realistic character motion sequences by temporally entangling the diffusion process with the motion's progression. It produces diverse and stable animations, avoiding common stitching artifacts and motion collapse, and supports interactive control.
Key insights
- The gradual denoising nature of Denoising Diffusion Probabilistic Models can be effectively adapted to the temporal dimension of sequential data, enabling continuous generation.
- A stationary motion buffer, combined with recursive frame output, allows for arbitrary long-term motion synthesis while inherently resolving stitching artifacts.
- Training with a blend of random and monotonic noise schedules is crucial for the model to learn both individual pose denoising and temporal coherence, leading to diverse and stable long-range motions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.