Wan Dancer

Wan-Dancer is a research model from Alibaba (HumanAIGC) for generating long dance videos directly from music. It addresses the roughly 20-second limit of typical diffusion video models by splitting generation into two stages: global keyframe planning using the full music track, and local temporal refinement between keyframes.

The paper describes time-mapped RoPE embeddings for dynamic frame-rate adaptation, an optical-flow-based loss for motion continuity, and motion-speed control for fast movements. It reports stable 720p/30fps output longer than one minute, with demo clips up to about 2 minutes 40 seconds, across five dance genres (Chinese classical, K-pop, street, tap, Latin). Generation can be conditioned on audio and text prompts.

Additional features shown on the project page include LoRA-based customization of specific choreography, outfit changes through edited keyframes, and manual motion keyframes for tight movement control. The page links to GitHub, ModelScope, and Hugging Face releases, and a YouTube creator reported the total download at about 85 GB, implying a high-end GPU is needed. The page notes that demo videos were post-processed after generation.