Article: Music generation controllable dance based on improved transformer model and style consistency Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.58 pp.69 - 84 Abstract: Music-driven dance generation can effectively improve the efficiency and popularity of artistic creation, but existing generation methods have problems such as insufficient dance-music correlation, poor stability of long sequence generation, and inconsistent styles. Therefore, a novel dance generation and completion framework that integrates improved transformer and style consistency control is proposed. This framework first constructs a bidirectional attention mechanism cross-modal generation model, enhances the correlation between dance and music through bidirectional interaction perception between music and action modalities, and adopts a planned sampling strategy to alleviate exposure bias in autoregressive generation. By extracting and integrating music features, key action features, and global dance style features, the completed dance segments ensure consistency in music synchronisation and overall style. Experiments showed that the generative model significantly outperformed mainstream comparison models in Frechet distance (25.7), beat coverage (59.7%), hit rate (52.4%), and diversity metrics. The complementary model achieved a style classification accuracy of 95.4% and a style retention rate of 90.2% in dance completion tasks. From this, the model proposed by the research can effectively improve the correlation and style consistency of generated dance, and promote the popularisation of art. Inderscience Publishers - linking academia, business and industry through research

Title: Music generation controllable dance based on improved transformer model and style consistency

Authors: Xiaoyun Cao; Yibo Wang

Addresses: School of Music, Guangdong University of Education, Guangzhou 510000, China ' School of Music, South China Normal University, Guangzhou 510000, China

Abstract: Music-driven dance generation can effectively improve the efficiency and popularity of artistic creation, but existing generation methods have problems such as insufficient dance-music correlation, poor stability of long sequence generation, and inconsistent styles. Therefore, a novel dance generation and completion framework that integrates improved transformer and style consistency control is proposed. This framework first constructs a bidirectional attention mechanism cross-modal generation model, enhances the correlation between dance and music through bidirectional interaction perception between music and action modalities, and adopts a planned sampling strategy to alleviate exposure bias in autoregressive generation. By extracting and integrating music features, key action features, and global dance style features, the completed dance segments ensure consistency in music synchronisation and overall style. Experiments showed that the generative model significantly outperformed mainstream comparison models in Frechet distance (25.7), beat coverage (59.7%), hit rate (52.4%), and diversity metrics. The complementary model achieved a style classification accuracy of 95.4% and a style retention rate of 90.2% in dance completion tasks. From this, the model proposed by the research can effectively improve the correlation and style consistency of generated dance, and promote the popularisation of art.

Keywords: transformer; bidirectional attention; BA; style consistency; music-driven dance generation; controllable dance.

DOI: 10.1504/IJICT.2026.153804

International Journal of Information and Communication Technology, 2026 Vol.27 No.58, pp.69 - 84

Received: 19 Jan 2026
Accepted: 28 Feb 2026
Published online: 26 May 2026 *

Title: Music generation controllable dance based on improved transformer model and style consistency

Keep up-to-date