MNET++: Music-Driven Pluralistic Dancing Toward Multiple Dance Genre Synthesis.

Area of Science:

• Artificial Intelligence
• Computer Graphics
• Computational Creativity

Background:

• Existing autoregressive networks for dance generation often produce repetitive and inferior dance sequences.
• A key limitation is the inability of current algorithms to generate diverse and non-repetitive dance motions from a given initial pose.

Purpose of the Study:

• To address the limitations of existing dance generation models, specifically repetitive patterns and lack of diversity.
• To propose a novel model architecture (MNET++) and training methodologies for enhanced dance synthesis.

Main Methods:

• Developed MNET++ incorporating a beat synchronizer for music rhythm adherence and a dance synthesizer for patch-based motion inference.
• Utilized adversarial learning with a transformer architecture to generate diverse dance sequences.
• Implemented a dance genre-aware latent representation for fine-grained user control across multiple domains.

Main Results:

• MNET++ successfully generates locally and globally consistent dance sequences synchronized with music beats.
• The model circumvents repetitive patterns, producing realistic and diverse dance motions.
• Qualitative and quantitative evaluations show superior performance compared to state-of-the-art methods in generating plausible dance sequences across multiple genres.

Conclusions:

• MNET++ represents a significant advancement in AI-driven dance generation, offering realistic, diverse, and controllable dance sequences.
• The proposed beat synchronizer and dance synthesizer modules effectively address key challenges in previous models.
• The genre-aware latent representation enhances user control and scalability for various dance domains.