Reproduce “Spatiotemporally Decoupled Autoregressive Diffusion Model for Human Motion Generation” 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
Chengqun Yang, Liang Xu, Yanping Li, Fulong Liu, Jingnan Gao, Weili Zeng, Yichao Yan
DeMoDiff introduces a spatiotemporally decoupled autoregressive diffusion model that generates high-fidelity 3D human motions from text descriptions, enabling fine-grained temporal and spatial editing while achieving an MPJPE of 9.4 mm and PA-MPJPE of 2.7 mm in motion reconstruction on HumanML3D.
Key insights
- Decoupling motion into a continuous 2D latent space per joint significantly improves reconstruction accuracy and detail retention.
- A combination of spatial-temporal, spatial, and temporal attention in a transformer effectively captures complex motion dynamics for generation.
- Bidirectional attention and masking in the diffusion framework enable intrinsic fine-grained temporal and spatial motion editing without requiring dedicated fine-tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.