Abstract
The generation of natural human motion interactions is a hot topic incomputer vision and computer animation. It is a challenging task due to thediversity of possible human motion interactions. Diffusion models, which havealready shown remarkable generative capabilities in other domains, are a goodcandidate for this task. In this paper, we introduce a novel bipartite graphdiffusion method (BiGraphDiff) to generate human motion interactions betweentwo persons. Specifically, bipartite node sets are constructed to model theinherent geometric constraints between skeleton nodes during interactions. Theinteraction graph diffusion model is transformer-based, combining somestate-of-the-art motion methods. We show that the proposed achieves newstate-of-the-art results on leading benchmarks for the human interactiongeneration task.