LongCat Avatar Vocal Model

LongCat_Video_SM_VocalModel loads the ONNX vocal-separation model that powers the pack's LongCat Avatar Vocal Extract node. It's the smallest node in the pack and entirely optional - if you only feed the avatar clean speech clips, you'll never touch it. But the moment you want the avatar to lip-sync to a song, a podcast with background music, or any recording where a voice shares space with instruments, this is the loader you reach for. It's the "model" half of the vocal chain: load a separator here, extract vocals with LongCat Avatar Vocal Extract.

The input and output

One input, one output:

• audio_encoder_vocal - a dropdown of .onnx files found in ComfyUI/models/longcat/, plus a none option. The supported model is Kim_Vocal_2.onnx.
• Output is an AUDIO_ENCODER that feeds LongCat Avatar Vocal Extract.

Selecting none is allowed - the node then loads a stub separator, which is fine if you only run clean-audio workflows. You only need a real model if you're actually separating vocals.

The model file

Grab Kim_Vocal_2.onnx from the UVR models repo and drop it in place:

# from huggingface.co/seanghay/uvr_models
mkdir -p ComfyUI/models/longcat
mv Kim_Vocal_2.onnx ComfyUI/models/longcat/

That's the entire install for this node's model. The Python dependencies it needs (onnxruntime, audio-separator) are already in the pack's default requirements.txt, so you don't install anything extra. Note this file is not in any auto-download manifest - it's a manual placement, and only needed when you use the vocal nodes.

Install

Pack install is shared across all nine nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI-LongCat-Avatar
cd ComfyUI-LongCat-Avatar
pip install -r requirements.txt

or ComfyUI Manager → search ComfyUI-LongCat-Avatar → restart. The vocal path also wants an ffmpeg executable visible on the PATH used to launch ComfyUI - pip install of imageio-ffmpeg doesn't satisfy the subprocess lookup, so sudo apt-get install ffmpeg (or the equivalent for your OS) is the reliable route.

Where people get burned

The dropdown being empty is the classic: the node lists .onnx files only from models/longcat/, and if the file is sitting in models/onnx or anywhere else, it won't show. Put it exactly in ComfyUI/models/longcat/. The second gotcha is scope - people install the whole vocal chain because a tutorial showed it, then wonder why the sampler ignores it. You only need this node if you're running LongCat Avatar Vocal Extract. Otherwise skip it; none keeps the graph honest. And if vocal extraction errors out with an ffmpeg complaint, that's the missing PATH executable, not the node.