XB_WanDancerCombo

Wan-Dancer-14B, Alibaba's music-to-dance model, needs three things before it can make a character move to a song: a slice of the audio, a set of audio features (onset envelope, MFCC, chroma, beats), and a text prompt describing the dance style. Upstream ComfyUI splits that across an audio slicer, a dance-audio encoder, and a prompt switcher. XB_WanDancerCombo collapses all three into one node: pick an audio file, set the clip times and dance style, and it hands you the audio features and the ready-made Chinese prompts.

The source is honest about its ancestry - the interface is "照抄 AudioSlicerV1" (copied from the pack's audio slicer), the feature extraction is "照抄 WanDancerEncodeAudio" (the official Wan Dancer encoder), and the prompt logic is the pack's XB_WanDanceSwitcher. So this is a convenience bundle, not new science: it slices the audio to your start/end time, computes the same mel-spectrogram/onset/MFCC/chroma/beat features the official encoder does, and generates the two prompts.

What you set

• audio - a file picker (mp3, wav, flac, ogg, m4a, ...) from your input directory; "none" is allowed.
• start_time / end_time (seconds) - which slice of the track to use.
• video_frames (default 149) - the total video frame count you plan to generate; the FPS is derived from the audio slice length vs. this count.
• audio_inject_scale (default 1) - how strongly the audio features drive the motion. Lower it if the dance gets too twitchy.
• dance_style - 古典舞 (classical), 韩舞 (K-pop), 街舞 (street/hip-hop), 拉丁舞 (Latin), 踢踏舞 (tap).
• motion_amplitude - 低 / 中等 / 高 / 最大 (low / medium / high / max).

Outputs: 音频 (AUDIO) - a bundle carrying the sliced waveform, the extracted audio_feature, the derived fps, and the inject scale; 整体描述提示词 (the full scene prompt: "一个人正在跳舞，舞蹈种类是X, 图像清晰程度高, 人物动作幅度Y"); and 舞蹈风格提示词 (the style + FPS string, e.g. "古典舞, 帧率是30fps。"). The audio bundle plugs into XB_WanDancerVideo's audio_encoder_output; the prompts go to a text encoder.

The dependencies - this one has teeth

Because it calls ComfyUI's official Wan Dancer module and torchaudio, it needs: torchaudio (pip install torchaudio) and scipy, plus a ComfyUI new enough to ship comfy_extras.nodes_wandancer - the node's own error says 0.28.0 or newer. If you cloned the pack on an older ComfyUI, this node will refuse loudly. The other thing to know: the prompts are Chinese, and the model was fine-tuned to respond to them, so resist "translating" them to English - the pack's whole Dancer workflow is built around these exact strings.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/WJLUOXIAO/XB_ToolBox.git
pip install torchaudio scipy

then restart, or install via ComfyUI Manager ("XB_ToolBox") and let it handle requirements. Wan-Dancer-14B weights are Apache 2.0 and ungated on HuggingFace.