Run Wan-Dancer-14B in ComfyUI

Hardware reality check

This is a 14-billion-parameter video model. As a rule of thumb for models in this class:

• 24 GB VRAM (RTX 3090/4090) — workable with quantized or FP8 variants and offloading
• 32–48 GB (RTX 5090, A6000, L40S) — comfortable at full 720p length
• Renting an A100/H100 by the hour on RunPod or Vast.ai is the usual shortcut if your local card is smaller

Exact requirements depend on the variant and quantization you pick — check the model card on Hugging Face for current numbers before downloading.

Setup steps

1. Install or update ComfyUI — Wan model support lands in core updates, so being current matters.
2. Download the Wan-Dancer-14B weights from the official Hugging Face repo and place them in your ComfyUI models directory (the model card specifies the exact subfolders for the diffusion model, text encoder and VAE).
3. Load the official workflow JSON from the Wan-AI release — it wires up the audio input, reference image and style prompt nodes for you.
4. Drop in your inputs: one vertical full-body photo, one audio track, one style prompt (K-pop, Chinese Classical, Street, Latin or Tap).
5. Queue it. First runs are slow (the model plans keyframes across the whole track before rendering) — a minute of 720p output takes serious compute even on big cards.

Tips that save time

• Start with a 15–20 second audio clip to validate your setup before committing to full-song renders.
• Reference photo quality dominates output quality: even lighting, full body visible, minimal occlusion.
• LoRA fine-tuning is officially supported — custom choreography styles are trainable if you have paired data.