One reference image + one music track → a beat-synced, full-body dance video (720p, 30 fps). Wan-Dancer is Tongyi Lab's open-weights dance specialist (Apache 2.0): a two-stage pipeline — a global model lays down choreography structure synced to the music's beat, then a local model runs a second pass refining motion detail. Styles out of the box: Chinese classical, K-pop, street, tap, Latin; it generalizes to described fusions.

I'm creating this as a taggable home for Wan-Dancer content on CivitAI, carrying Comfy-Org's official workflow template, unmodified (MIT, maintained by them - source / tutorial), plus everything you need to run it. The model weights are NOT re-hosted; Hugging Face serves them canonically (links below).

Full disclosure: I don't generate through ComfyUI. My posts to this workflow come from an equivalent native pipeline implementing the same recipe. This is the official workflow and is expected to work as-is; if it breaks, comment and I'll re-sync it from upstream.

Get the pieces

• fp8 stage models (consumer VRAM): Comfy-Org/Wan-Dancer → diffusion_models/

• bf16 originals (big cards): Wan-AI/Wan-Dancer-14B → diffusion_models/

• Lightning LoRA (rank64): on CivitAI / Kijai/WanVideo_comfy → loras/

• umT5 text encoder → text_encoders/

• CLIP-vision H → clip_vision/

• Wan 2.1 VAE → vae/

• All above per the tutorial

The recipe

• Lightning (default, ~8× faster): rank64 LoRA at strength 3.0 on the global model and 1.03 on the local; 6 steps per stage, CFG 1, sigma shift 5 (ModelSamplingSD3). Each stage is a full pass! Global drafts the clip, local re-samples it conditioned on the draft.

• Full schedule (no LoRA): 48 steps global / 24 local, CFG 5, shift 5. Will take hours on a single big GPU.

• 720×1280 (or 1280×720).

• WAN Dancer is optimized for minute-scale dance videos. Generating a 5 second clip is possible, but anecdotally, the intensity of the dance seems to get lower once the clip length falls below 15. Adjust your intensity to compensate.

Prompting

In the workflow, you don't type the prompt at all: the Dance Style and Motion Amplitude dropdowns assemble it. If you want to expand or to build on this, the prompts are visible in the source. They are short and simple prompts in the Chinese language.

If your character doesn't dance

I'm far from an expert in this model or workflow, but I encountered the no-dancing issue on many sample creation attempts. I was able to confirm this was caused by the start image I was using. Although the feet were clearly in the image, I think the character was too close to the camera.

Generally, the way to troubleshoot is to change just one thing at a time, making it more and more like some successful run you're aware of, until you're doing exactly what that successful run did. Hopefully, you don't get all the way to running the whole model and 48 steps.

Credits & licenses

Model: Tongyi Lab / Wan team, Apache 2.0.
Lightning distill: lightx2v, Apache 2.0; rank64 extraction: Kijai.
Workflow template: Comfy-Org, MIT (notice included with the file). All credit for the model and workflow belongs to them.