Wan-Dancer¶

Development line: project:wan-dancer · thread wan-dancer-development
Last event: 2026-07-14 · 1 dated since 2026-07-14 · Researched: 2026-09-06 · confidence: medium

What it is¶

Wan-Dancer is a 14B music-to-dance video system for creators who need a long choreographed clip from a portrait, music track and style prompt rather than manually stitching short clips.

• Global keyframe planning from the complete track.
• Local refinement for temporal continuity.
• Style prompts for Chinese Classical, K-Pop, Street, Tap and Latin dance.

Claimed output runs at 720p and 30fps for more than one minute. The authors tested their code on 8 × NVIDIA A800 80GB GPUs. We use the author pipeline for controlled long-form work, not as a validated single-GPU or raw-demo workflow.

Development line¶

• 2026-07-14 — Wan-Dancer project resources were publicly linked. On 2026-07-14, public links appeared for the Wan-Dancer project website, source repository, Hugging Face model page, and ModelScope studio. This marked a distribution milestone for the project. The links do not establish technical claims, version details, or original announcement wording.

What changed¶

Wan-Dancer — development line.

• 2026-05-09 — ComfyUI merged core WanDancer support before the public model release; this reflects implementation history rather than public artifact availability.
• 2026-07-10 — arXiv v1 published the hierarchical music-to-dance method.
• 2026-07-13 — Wan-AI announced the release of Wan-Dancer-14B weights and inference code.
• 2026-07-14 — The official repository pushed an update-links commit, and a ModelScope Community post documented a Studio portrait-and-music route.
• 2026-07-17 — arXiv published v3 without describing behavioral or weight changes in its metadata.

How to use this¶

As of 2026-07-14, check the linked project page for source code and official model distributions, then verify access, licensing, and instructions before adopting it.

1. Clone Wan-Video/Wan-Dancer, create the virtual environment and install the pinned dependency set from the README. The reference platform is Ubuntu 22.04, Python 3.10.14 and CUDA 12.4. — https://github.com/Wan-Video/Wan-Dancer
2. Download Wan-AI/Wan-Dancer-14B with huggingface-cli into a controlled local model directory. — https://huggingface.co/Wan-AI/Wan-Dancer-14B
3. Prepare image_path, music_path and a global prompt_path; prompt families cover Chinese Classical, K-Pop, Street, Tap and Latin dance. — https://huggingface.co/Wan-AI/Wan-Dancer-14B
4. Run gen_video_global.sh and save the generated global video with its seed and settings. — https://github.com/Wan-Video/Wan-Dancer
5. Pass that result as global_video_path to gen_video_local.sh with the matching local prompt so local refinement has its base video. — https://huggingface.co/Wan-AI/Wan-Dancer-14B
6. For ComfyUI, update ComfyUI, open its Wan-Dancer template, and place global/local models, LoRA, text encoder, CLIP-Vision model and VAE in named template folders. This template changes upstream and is not locally verified here. — https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/templates/video_wan_dancer.json

Best practices¶

• Fix the seed across comparisons so runs remain reproducible. — https://huggingface.co/Wan-AI/Wan-Dancer-14B
• Treat global and local generation as one pipeline, and never replace the required global_video_path with an unrelated clip. — https://huggingface.co/Wan-AI/Wan-Dancer-14B
• Raise inference steps for longer videos; the authors suggest 48 steps rather than treating the 24-step local example as universal. — https://huggingface.co/Wan-AI/Wan-Dancer-14B
• Do not evaluate raw quality against the project gallery, because displayed videos underwent post-processing. — https://humanaigc.github.io/wan-dancer-project/
• Start at the 5-second default in the ComfyUI template and monitor VRAM, because sampling 149 frames at once demands heavy memory. — https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/templates/video_wan_dancer.json
• Do not treat the authors' 8 × A800 80GB test system as a single-GPU minimum, because it serves as a reference environment rather than a hardware floor. — https://github.com/Wan-Video/Wan-Dancer

Superseded by this¶

• Treating 2026-07-14 as the first public release date for Wan-Dancer weights and inference code; the official model card dates that release to 2026-07-13.

Still unknown¶

• The direct ModelScope Studio page did not expose live controls, leaving its current fields, defaults, access policy, and limits unverified.
• No author source published a single-GPU minimum VRAM, runtime, or end-to-end reproduction; the only reported setup is 8 × A800 80GB.
• Displayed videos had post-processing and lack independent reproduction, so reported duration and quality are not production guarantees.
• Checkpoint lineage and compatibility between Wan-Dancer-14B and the wan2.2_dancer assets in the ComfyUI template remain unverified.
• ComfyUI support status remains unresolved: the official model card marks integration incomplete, while ComfyUI merged a working template.
• Checked sources show no official Wan-Dancer feature or weight update after 2026-07-17, though unindexed releases may exist.

Sources¶

Agent brief¶

• Subject: project:wan-dancer, thread wan-dancer-development, 1 dated events 2026-07-14 → 2026-07-14.
• Practical note: As of 2026-07-14, practitioners can use the linked project page to find Wan-Dancer's source and official model-distribution resources, then verify current access, licensing, and usage instructions before adopting it.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.