A hierarchical generative framework for the computational re-animation of Dunhuang dance

The transmission of contemporary Dunhuang dance (DHD) presents a distinctive challenge: its continuation depends critically on domain experts who, drawing from fragmentary and static cave-mural poses, creatively reconstruct coherent dynamic performances. This tacit creative process is inherently difficult to capture objectively or to pass on through conventional documentation. To address this gap, we introduce a generative framework for Computational Re-animation that, to our knowledge, is the first to computationally model this expert creative workflow. Our approach is founded on a newly curated, large-scale, high-fidelity optical motion-capture dataset of DHD and implemented via a hierarchical generative architecture. At the top level, a duration prediction network (DPN) models the choreographer’s implicit sense of rhythm and forecasts the temporal extent of motion segments; at the lower level, a physics-informed diffusion model (DDPM) synthesises motion sequences that adhere to the distinctive “stylistic grammar” of DHD. Extensive quantitative and qualitative evaluations demonstrate that our framework improves realism, rhythmic coherence, and diversity over prior methods, while externalising expert tacit creative knowledge into a computable model for active, exploratory re-animation. It also provides a reproducible substrate for humanities and social-science studies of mural-driven heritage activation, transmission and public engagement through dance.