SD-T2LM: Long-Sequence Human Motion Generation Based on Dynamic SLERP Interpolation and Dynamic Mask Mechanism

Abstract

Text-driven human motion generation aims to synthesize coherent, natural human actions from natural-language descriptions. While current techniques are effective for short sequences, they often suffer from abrupt transitions, detail loss, and temporal misalignment when extended to long sequences. We present SD-T2LM, a unified framework that couples transition optimization, detail enhancement, and temporal alignment to produce natural, detail-rich long motion sequences. First, a dynamic SLERP interpolation selects transition length conditioned on start/end pose similarity to suppress discontinuities while preserving physical plausibility. Second, a dynamic masking mechanism progressively reduces the mask ratio from 60% to 30% during training, enabling staged learning-global structure early, fine-grained details later. Third, a Temporal Flexible Alignment loss based on Soft-DTW aligns generated and ground-truth rhythms without enforcing point-wise synchrony. On the HumanML3D dataset and KIT-ML dataset, SD-T2LM showed significant improvements in FID and diversity compared to the baseline model, while reducing transition distance and jerk.