Logo
Wan Dancer
Loading account

Pollo AI: Free All-in-One AI Image and Video Generator

Try for Free

Music to dance · 14B · open model

Wan-Dancer — minute-scale dance video from music

Turn a portrait, a track, and a style prompt into rhythm-synced dance video. Wan-Dancer-14B keeps identity and choreography coherent past the usual ~20s diffusion limit — stable 720p / 30fps clips over one minute.

Try the playgroundRead the paper
Model
Wan-Dancer-14B
Output
720p / 30fps
Try in browser
HF Spaces

Why Wan-Dancer

Built for long, on-beat dance — not short loops

Most video diffusion models break down after ~20 seconds. Wan-Dancer is a hierarchical music-to-dance framework from Tongyi Lab / Wan-AI that plans globally, then refines locally.

  • Minute-scale coherence

    Full-track musical context drives global keyframe planning so motion structure stays consistent beyond short clip length.

  • Identity-consistent motion

    Reference image conditioning keeps the subject stable while rhythm and style follow the audio and text prompts.

  • Five dance genres

    Demonstrated across Chinese classical, street, K-pop, Latin, and tap — conditioned on audio and style prompts.

  • Open weights (Apache 2.0)

    Model weights on Hugging Face and ModelScope, with inference code on GitHub and community browser demos.

How it works

Two-stage hierarchical pipeline

Wan-Dancer decouples long-horizon planning from high-resolution refinement so choreography stays musically structured without identity collapse.

  1. 1

    Global keyframe video

    Plan full-track choreography from music + reference image + style prompt, using the entire musical context.

  2. 2

    Local high-res refinement

    Condition on the global plan to produce the final 720p/30fps clip with tighter rhythm and detail.

  3. 3

    Time-mapped RoPE & flow losses

    Dynamic frame-rate alignment and optical-flow continuity reduce temporal drift on long sequences.

  4. 4

    Motion-speed control

    Preserve fidelity during fast moves instead of smearing or collapsing into repetitive patterns.

How to use it

A simple starting recipe

Try ideas in the browser playground first, then run the open model locally or in ComfyUI when you need full control.

  • 1. Prepare inputs

    A vertical portrait (identity), a music clip, and a short dance style / scene prompt.

  • 2. Try the playground

    Open /playground and use the running ZeroGPU Space to generate a short demo without installing weights.

  • 3. Run open weights

    Download Wan-AI/Wan-Dancer-14B from Hugging Face or ModelScope and follow the official GitHub inference code.

  • 4. ComfyUI workflows

    Community fp8 / GGUF packs and ComfyUI graphs are available for lower-VRAM local runs.

Playground

Try Wan-Dancer in the browser

Switch tabs between community Hugging Face Spaces that load Wan-Dancer-14B. No local GPU install required to start.

Spaces are third-party. Availability, queue times, and loaded weights can change upstream. Some demos may be paused or non-generative.

Launch playground

Resources

Where to go next

Browser playground

Embed community HF Spaces for Wan-Dancer music-to-dance generation in one place.

Open playground

Paper (arXiv)

Wan-Dancer: A Hierarchical Framework for Minute-scale Coherent Music-to-Dance Generation.

Read paper

Model & code

Weights on Hugging Face / ModelScope and inference code on GitHub (Wan-Video/Wan-Dancer).

Open Hugging Face