Li Ronghui ; Zhang Hongwen ; Zhang Yachao ; Zhang Yuxiang ; Zhang Youliang ; Guo Jie ; Zhang Yan ; Li Xiu ; Liu Yebin

Abstract

We propose Lodge++, a choreography framework to generate high-quality,ultra-long, and vivid dances given the music and desired genre. To handle thechallenges in computational efficiency, the learning of complex and vividglobal choreography patterns, and the physical quality of local dancemovements, Lodge++ adopts a two-stage strategy to produce dances from coarse tofine. In the first stage, a global choreography network is designed to generatecoarse-grained dance primitives that capture complex global choreographypatterns. In the second stage, guided by these dance primitives, aprimitive-based dance diffusion model is proposed to further generatehigh-quality, long-sequence dances in parallel, faithfully adhering to thecomplex choreography patterns. Additionally, to improve the physicalplausibility, Lodge++ employs a penetration guidance module to resolvecharacter self-penetration, a foot refinement module to optimize foot-groundcontact, and a multi-genre discriminator to maintain genre consistencythroughout the dance. Lodge++ is validated by extensive experiments, which showthat our method can rapidly generate ultra-long dances suitable for variousdance genres, ensuring well-organized global choreography patterns andhigh-quality local motion.