TransMotion: Text-Guided Transformation Learning for Composed Human Motion Retrieval

Text-to-motion retrieval systems enable users to find 3D human motion using natural-language descriptions. However, conveying highly specific intent with a single text query is often difficult, and users may wish to find a motion that preserves the core semantics of an existing motion while reflecting specific modifications. To address these scenarios, we introduce Composed Human Motion Retrieval (CHMR), a new task that aims to retrieve a target motion from a database using a composed query that combines a reference motion with a textual modification instruction. We propose TransMotion, a framework designed for the CHMR task. TransMotion supervises the motion representation through two complementary objectives, yielding invariant and adaptive components. The invariant component preserves the core semantic identity between the reference and target motions, while the adaptive component models the dynamic transformation specified by the text instruction. Through extensive experiments, we determine that a contrastive learning objective paired with a cross-attention-based transformation predictor yields the most effective configuration, which we attribute to the attention mechanism’s capacity to model complex text–motion relationships. We validate our approach on two independently constructed datasets, MotionFix and STANCE, demonstrating superior performance over native composed-retrieval baselines and strong baselines adapted from motion editing and standard retrieval domains.