AnyTop: Character Animation Diffusion with Any Topology

AnyTop is a transformer-based diffusion model for generating diverse character motions using skeletal structure input. It features topology integration, semantic joint correspondence learning, and a latent space for downstream tasks. AnyTop excels in generalizing with minimal training examples, supports unseen skeletons, and offers an informative latent space. Its webpage includes videos and code. The model demonstrates three forms of generalization: In-skeleton, Cross-skeleton, and Unseen-skeleton. It enables motion generation for various characters and skeletons, including non-homeomorphic ones, not seen during training. Beyond generative capabilities, it uses Diffusion Features for downstream applications like unsupervised correlation, temporal segmentation, and motion editing.