GestureFAR: Streaming Co-Speech Gesture Generation with Flow Autoregression

arXiv cs.GR details GestureFAR, a streaming co-speech gesture generation framework built for embodied conversational agents that need motion while speech is still unfolding. The key shift is moving away from discrete motion tokens and toward causal continuous latents, which helps preserve expressiveness without breaking online generation.

GestureFAR pairs a transformer backbone for streaming audio-motion context with a per-token flow-matching head that samples the next latent from a continuous distribution. That matters for developers building real-time avatars, NPCs, or virtual assistants, because gesture quality often drops when systems are forced through compact codebooks or delayed batch inference.

The paper also introduces a head-only flow distillation step that freezes the causal backbone and collapses the multi-step flow head into a single network evaluation. In practice, that is meant to remove the main latency bottleneck while keeping token-causal behavior intact, which is the hard part for live interaction systems.

On BEAT2, the authors say GestureFAR improves the quality-latency trade-off over other streaming-capable methods, while still supporting real-time generation. The exact production targets and runtime costs for a game engine integration haven't been disclosed, but the approach is relevant anywhere speech-driven animation needs to feel immediate and natural.