TPPMG: Temporal Planning-driven Progressive Motion Generation

Abstract

Most text-to-motion models adopt a fixed-length sampling paradigm, treating time as a hyperparameter rather than a decision variable inferred from semantics. This leads to a systematic temporal mismatch between text and generated motion: short prompts suffer from duration trailing, while long prompts exhibit stage-wise semantic collapse. To address this issue, we propose TPPMG (Temporal Planning-driven Progressive Motion Generation), a temporally planned progressive framework that follows the pipeline semantics, temporal structure, and motion realization. TPPMG has two key components: (1) a Temporal Planner, which uses an LLM with a duration calibrator to infer stage boundaries and durations from text, yielding an explicit semantics-to-temporal structure mapping; and (2) a Progressive Diffusion Generator, which introduces heterogeneous noise scheduling within fixed windows and a sliding-window mechanism at inference to resolve the mismatch between fixed-length diffusion training and variable-length sampling, thereby generating variable-length motion sequences that match the planned temporal structure. We further build three benchmarks: HumanML3D-Short for short texts, and two multi-stage benchmarks HumanML3D-Concat and BABEL-Concat, along with a comprehensive evaluation suite. Experiments show that TPPMG effectively suppresses duration trailing for short texts and simultaneously improves stage-structure accuracy, segment-level semantic consistency, and boundary continuity for long texts.