Lodge++: High-Quality and Long Dance Generation With Robust Choreography Patterns.

Area of Science:

• Computer Graphics and Animation
• Artificial Intelligence
• Computational Creativity

Background:

• Generating long, high-quality, and physically plausible dances computationally is challenging.
• Existing methods struggle with global choreography patterns and computational efficiency for ultra-long sequences.

Purpose of the Study:

• To introduce Lodge++, a novel choreography framework for generating ultra-long, high-quality dances.
• To address challenges in computational efficiency, global choreography, and physical plausibility in dance generation.

Main Methods:

• A two-stage strategy: a global choreography network for dance primitives, followed by a primitive-based dance diffusion model.
• Incorporation of a penetration guidance module, a foot refinement module, and a multi-genre discriminator for physical realism and genre consistency.

Main Results:

• Lodge++ rapidly generates ultra-long dances across various genres with well-organized global choreography.
• The framework ensures high-quality local motion and adherence to complex choreography patterns.
• Experimental validation confirms the method's effectiveness in producing realistic and genre-consistent dances.

Conclusions:

• Lodge++ offers an efficient and effective solution for generating complex, ultra-long dances.
• The proposed framework advances the state-of-the-art in computational choreography and motion synthesis.