Reproduce “Executing your Commands via Motion Diffusion in Latent Space” 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
Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, Jingyi Yu, Gang Yu
Researchers from Tencent PCG and Fudan University developed Motion Latent-based Diffusion (MLD), a model that generates high-quality and diverse human motion sequences from various commands by applying a diffusion process in a learned low-dimensional latent space. MLD achieves state-of-the-art performance while reducing inference time by two orders of magnitude compared to previous raw motion diffusion models.
Key insights
- Conducting the diffusion process within a low-dimensional, learned latent space drastically improves computational efficiency and robustness to noise in raw motion data.
- A transformer-based Variational AutoEncoder (VAE) with long skip connections effectively learns a robust and information-dense latent representation of human motion.
- Concatenation is an effective method for injecting conditional information (text, action) into the latent diffusion model's denoiser, outperforming cross-attention.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.