A new co-speech gesture system aims to make digital humans move more naturally while cutting inference from dozens of steps to a faster flow-matching pipeline. GestureLSM models body-region interactions with spatial and temporal attention, then adds latent shortcut learning to improve quality. The result is state-of-the-art performance on BEAT2 with a practical speedup for real-time embodied agents.
arXiv cs.GR
arXiv cs.GR
1% match
With your permission, GameDev.net uses analytics cookies to understand how people use the platform. You can accept analytics or continue with necessary cookies only. Learn more
