DC-Motion: Decoupling Semantics and Details via Discrete-Continuous Tokens for Human Motion Generation
Researchers have developed DC-Motion, a novel framework for generating human motion from text. This approach decouples semantic meaning from fine-grained physical details using a combination of discrete and continuous tokens. A Discrete-Continuous VAE first separates motion into discrete semantic tokens and continuous dynamic residuals. Subsequently, a masked autoregressive model interprets text to predict the discrete structure, while a diffusion model reconstructs the continuous physical details. Experiments show DC-Motion excels in following complex instructions and achieves state-of-the-art performance on datasets like HumanML3D and KIT-ML for both motion realism and text alignment. AI