L-DANCE uses a prompt-tuned Llama 3.2 11B model as a strategic planner to derive an executable SelectorSpec from natural language instructions. The framework then generates multiple dance candidates with a frozen diffusion backbone and ranks them using a LoRA-tuned Llama 3.2 1B dynamics-aware selector.
L-DANCE translates natural language dance instructions into an executable planning specification and selects the motion candidate that best matches the intended dynamics.
Prompt: "Genre: K-POP. Generate a dance that feels calm and grounded overall, but still includes confident and powerful accents at key musical moments. The motion should avoid being overly chaotic while preserving strong visual impact."
Strategy: weighted ranking to balance dynamic intensity and temporal stability.
Prompt: "Genre: Hiphop. Generate a soft and graceful dance with gentle transitions, curved trajectories, and smooth upper-body flow. The choreography should feel light and expressive rather than sharp or aggressive."
Strategy: min_jerk to favor smoother and more continuous motion trajectories.
Prompt: "Genre: POP. Generate a smooth and flowing dance sequence with continuous transitions between poses. The movement should feel stable, polished, and visually connected across the full phrase of the music."
Strategy: min_jerk to encourage continuous and non-abrupt temporal motion.
Prompt: "Genre: POP. Generate a jazzy dance with rhythmic groove, playful accents, and syncopated energy. The motion should feel lively and musical, while remaining coherent and physically stable over time."
Strategy: weighted ranking to preserve rhythmic energy while maintaining temporal stability.
Given the same music input, L-DANCE generates distinct choreographic textures depending on the instruction: high-energy motion with stronger accents (left, red) versus softer motion with smoother trajectories (right, green).
Quantitative results show that L-DANCE improves instruction alignment, physical naturalness, and choreographic expressiveness over the single-pass baseline.
| Model | IA ↑ | PN ↑ | CE ↑ | Preference (%) ↑ |
|---|---|---|---|---|
| Baseline (EDGE) | 3.21 ± 0.45 | 3.42 ± 0.38 | 3.15 ± 0.52 | 40.5% |
| L-DANCE (Ours) | 4.12 ± 0.31 | 3.95 ± 0.25 | 4.38 ± 0.41 | 59.5% |
@article{LDance2026,
title={L-DANCE: LLM-guided Dynamics-Aware Natural Dance Generation},
author={Anonymous Authors},
journal={Under Review},
year={2026}
}