GestureLSM: Latent Shortcut based Co-Speech Gesture Generation with Spatial-Temporal Modeling

GestureLSM tackles two pain points that have held back speech-driven character animation: awkward full-body motion and slow generation. Instead of treating body parts in isolation, it tokenizes regions and uses spatial-temporal attention to model how hands, legs, and torso influence one another, which is meant to produce more coherent gestures.

The second half of the system swaps in flow matching to reduce sampling cost, then layers on latent shortcut learning and beta-distribution timestamp sampling to recover quality that simpler flow-matching baselines tend to lose. That combination is aimed at making gesture synthesis both cleaner and cheaper to run, which matters if the output needs to be generated interactively rather than as offline animation.

The paper reports state-of-the-art results on BEAT2 and says inference is significantly faster than existing approaches. For game teams building digital humans, NPC presenters, or embodied agents, that points to a path toward speech-reactive motion that is less disjointed and more feasible in real time. The exact integration cost for production pipelines still depends on the runtime stack, but the direction is clear: better motion quality without the usual diffusion-style latency tax.