This paper presents a novel method, Discrete Tokens to Continuous Motion via Rectified Flow Decoding (DisCoRD). Designed to address the differences between discrete and continuous motion representations, it utilizes rectified flow to decode discrete motion tokens into a continuous raw motion space. To address the limited expressiveness and frame-level noise artifacts of existing discrete generation methods, as well as the difficulty of continuous approaches in complying with conditional signals, we structure token decoding as a conditional generation task to capture subtle motions and generate smoother, more natural motion. We enhance naturalness while maintaining fidelity to conditional signals across a variety of settings, achieving state-of-the-art performance (FID 0.032 and 0.169, respectively) on the HumanML3D and KIT-ML datasets.
