Motion Puzzle: Arbitrary Motion Style Transfer by Body Part

This paper presents Motion Puzzle, a novel motion style transfer network that advances the state-of-the-art in several important respects. The Puzzle is first can control of individual body parts, allowing for local editing and significantly increasing range stylized motions. Designed to keep human's kinematic structure, our framework extracts features from multiple motions different parts transfers them locally target parts. Another major advantage it both global traits by integrating adaptive instance normalization attention modules while keeping skeleton topology. Thus, capture styles exhibited dynamic movements, such as flapping staggering, better than previous work. In addition, allows arbitrary without datasets with labeling or pairing, making many publicly available training. Our be easily integrated generation frameworks create applications, real-time transfer. We demonstrate advantages number examples comparisons