Reproduce “AMD: Autoregressive Motion Diffusion” 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
Bo Han, Hao Peng, Minjing Dong, Yi Ren, Yixuan Shen, Chang Xu
The Autoregressive Motion Diffusion (AMD) model synthesizes high-fidelity 3D human motion from complex, long text or audio inputs by combining autoregressive generation with diffusion processes. It also introduces HumanLong3D, a new dataset for multi-stage motions with extensive text descriptions, and HumanMusic for audio-to-motion tasks. AMD achieves an FID of 0.225 on HumanLong3D, outperforming previous state-of-the-art models, and demonstrates competitive performance in music-to-dance generation.
Key insights
- Combining an autoregressive framework with a denoising diffusion probabilistic model (DDPM) effectively handles long-term dependencies and variable motion durations for complex conditional generation.
- A context encoder that consolidates current and previous text prompts, alongside prior generated motion segments, provides crucial historical information for ensuring coherent motion synthesis.
- Integrating explicit physical constraints into the generative process is critical for producing not just statistically plausible but also physically realistic and natural-looking human movements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.