Subscribe

Wan-Dancer-14B

Open 14B model that turns music into minute-long, 720p/30fps dance video using keyframe planning plus local refinement.

Two-stage hierarchical design with global keyframe planning from the full track, then local temporal refinement, plus time-mapped RoPE for rhythm alignment and an optical-flow loss for motion continuity. Handles five dance genres from audio and text prompts. Apache 2.0.

Date
Friday 10 July 2026
Lab
Alibaba (Qwen)
Kind
open-weights
Access
open weights

Figures

MeasureValueMeasured by
Output lengthover 1 minute at 720p/30fps
Per the paper abstract (company claim); prior diffusion models usually fail beyond about 20 s
company

The date is when the Hugging Face repo was created, 2026-07-10. The arXiv paper 2607.09581 reached its third version on 2026-07-17.

Sources

  1. github.com/Wan-Video/Wan-Dancer
  2. arxiv.org/abs/2607.09581
  3. huggingface.co/Wan-AI/Wan-Dancer-14B

This record was checked against its sources on 6 October 2026. How we check

Related