Our method optimizes a fused semi-unbalanced Gromov-Wasserstein objective to align motion. It alternates between metric-alignment and patch blending to produce high-fidelity results.
Fig.2: Intuitive concept of MAMM
Fig.3: FSUGW objective and alternating optimization
Control periodicity and style via 1D waveforms (sine, square, center-alternating) to generate diverse motion.
Draw 2D curves to specify motion trajectories. Soft and hard keyframes allow fine-grained control.
Align motions using segmentation labels as one-hot vectors—no precomputed annotations required.
Cross-skeletal alignment enables direct mapping between different motion styles and skeletons.
Use MFCC features from music and speech to drive dance and gesture synthesis without training.