Satoshi Ikehata

Retrieving and Refining Winning Noise Tickets for Diffusion-Based Motion Generation

Sakuya Ota, Qing Yu, Kent Fujiwara, Satoshi Ikehata, Ikuro Sato

Diffusion-based text-to-motion models synthesize realistic human motions but often exhibit semantic drift from the input text. Motion is inherently temporal, especially in compositional and long-duration sequences that require semantic consistency across multiple action segments…Read more