MotionMERGE unifies fine-grained human motion editing and reasoning in one LLM framework

MotionMERGE is a new unified framework from researchers at multiple institutions that tackles the fundamental granularity gap in human motion-language models. Existing systems operate at a coarse level, lacking the fine-grained control over specific body parts and temporal sequences needed for animation or interactive applications. MotionMERGE addresses this by explicitly modeling motion at both part-level and temporal granularity within a single large language model (LLM), endowing it with robust priors for precise, localized control.

The framework introduces a novel Reasoning-Aware Granularity-Synergy pre-training strategy that jointly supervises cross-granularity alignment, temporal grounding, localized alignment, motion coherency, and motion-grounded chain-of-thought (CoT) reasoning. To support this, the team curated MotionFineEdit, a large-scale dataset containing 837K atomic triplets and 144K complex triplets — the first to include fine-grained spatio-temporal corrective instructions and motion-grounded CoT annotations. Extensive experiments demonstrate MotionMERGE's superior performance in motion generation, understanding, and editing, as well as compelling zero-shot generalization to other complex motion tasks, marking a significant step toward human-like motion interaction.