Refine Synthesized Action-Conditioned 3D Human Motion with Unsupervised Learned Motion Concepts

ICLR 2026 Conference Submission6573 Authors

Keywords: Action to Motion, Refinement, Motion Concepts

TL;DR: We propose the first motion refinement framework to the task of human motion synthesis utilizing unsupervisedly learned motion concepts to recognize and refine flawed frames in action-conditioned synthetic motion sequences hierarchically.

Abstract: As a fundamental aspect of human motion understanding, numerous efforts have been devoted to 3D human motion synthesis and tremendous progress has been made in recent years. Nevertheless, the synthesis of natural and seamless human motions still poses challenges, as it is inevitable for flawed frames to occur within the generated sequences. In light of this, there is a substantial demand for a refinement algorithm, an area that has received limited attention in previous research. In this work, we present motion concepts, which are unsupervisedly learned from a set of real motion sequences, to capture the common and regular patterns in human actions. By leveraging motion concepts, we propose a three-step framework to recognize and refine the flawed frames in an action-conditioned motion sequence hierarchically. Exhaustive experiments conducted on two widely-used benchmarks with four representative motion synthesis approaches, demonstrate that our refinement framework significantly elevates the performance of existing approaches, by improving the realism of synthesized motions while simultaneously enhancing their diversity and multimodality. Our code will be made publicly available.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 6573