What is Wan-Dancer-14B?

Why it's different from other video models

Most AI video models generate frames in sequence, and small errors compound: after roughly twenty seconds the dancer's face morphs, the outfit changes color, the motion loses the beat. Wan-Dancer takes a different route — it plans before it renders.

• Global stage:the model listens to the full track first and lays out the entire routine as a sequence of keyframes matched to the music's structure.
• Local stage: it then refines the motion between those keyframes frame by frame, so detail stays sharp while the long-range plan holds everything together.

That hierarchical design is why a Wan-Dancer video still shows the same person, on beat, past the one-minute mark.

What you need to generate a video

• One reference photo — a vertical, full-body shot works best
• A music or audio file
• A short text prompt choosing the dance style: K-pop, Chinese Classical, Street, Latin, or Tap

No motion capture, no driving video, no rigging. The choreography itself is generated from the music.

License and availability

The model weights, inference code, ComfyUI integration and LoRA fine-tuning tooling are all released under Apache-2.0 — free for commercial use. You can download everything from the official Hugging Face repository or read the research paper. If you have a GPU, our ComfyUI guide walks through running it locally. There is no official hosted API yet — see the current API status.