HyDance: A Novel Hybrid Dance Generation Network with temporal and frequency features
Keywords: Diffusion Models，Motion Generation
Abstract: We propose HyDance, a diffusion network utilizing both the temporal and frequency-domain representations of dance motion sequences for music-driven dance motion generation. Existing dance generation methods primarily use temporal domain representations of dance motion in their networks, which often results in the network losing the sfrequency-domain characteristics of the dance. This manifests in overly smooth generated dance motion sequences, resulting in dance movements that lack dynamism. From an aesthetic perspective, such overly smooth movements are perceived as lacking expressiveness and the sense of power. To address this issue, we designed HyDance, which incorporates independent temporal feature encoders and frequency-domain feature encoders. The model employs a shared-weight hybrid feature encoder, enabling the complementary extraction of motion information from both domains. By introducing compact frequency-domain features into the dance generation framework, our method mitigates the oversmoothing problem in generated dance motion sequences and achieves improved spatial and temporal alignment in the generation results. Experiments show that our method generates more expressive dance movements than existing methods and achieves better alignment with the music beats.
Supplementary Material: zip
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5560
Loading