This article presents the EnchantDance framework, a state-of-the-art method for music-driven dance generation. The framework addresses the challenges of high diversity in dance movements, lack of large-scale music-dance datasets, and maintaining a consistent dance style. The EnchantDance first constructs a strong dance latent space, then trains a diffusion model. To bridge the data gap, a large-scale music-dance dataset, ChoreoSpectrum3D, is created. A music genre prediction network is also incorporated to enhance the consistency between music genre and dance style. The framework has demonstrated high performance in dance quality, diversity, and consistency.
Publication date: 29 Dec 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2312.15946
