Abstract:

Full-body co-speech gesture generation significantly enhances the interactivity of virtual digital humans, requiring generated gestures to not only align accurately with speech but also demonstrate realistic full-body dynamics. To address limitations of existing methods—Transformer-based approaches often overlook temporal features of action sequences, while diffusion model-based ones inadequately capture spatial correlations between body parts, a full-body action generation method integrating diffusion models, Mamba, and attention mechanisms is proposed.We introduce the spatial self-attention-temporal state space model (STMamba Layer) as the core of denoising network to extract inter-part spatial features and intra-part temporal features, thus enhancing action quality and diversity. Body motion sequences are modeled in two dimensions: spatially, rotational relative positional encoding and self-attention capture spatial correlations among body joint points; Mamba captures intra-part temporal dynamics in action sequences to boost continuity. Experiments and evaluations on the large-scale audio-text-action dataset BEAT2 demonstrate that the proposed method outperforms state-of-the-art approaches in both fidelity and diversity, while maintaining competitive inference speed despite performance gains.

Key words: human Avatar, full-body co-speech gesture generation, conditional diffusion model, Transformer, Mamba