Dawei Guan, Di Yang, Chengjie Jin, Jiangtao Wang

Researchers propose FlowCoMotion, a text-to-motion generation framework that combines continuous and discrete motion representations to address limitations in existing approaches. The method uses token-latent coupling to preserve both semantic content and fine-grained motion details, applying multi-view distillation to the continuous latent space while using discrete temporal quantization for semantic extraction. A velocity field predicted from text conditions is then integrated via ODE solver to generate target motions. The approach achieves competitive results on HumanML3D and SnapMoGen benchmarks.