audio-separator-cli-node

Audio Separator CLI — ComfyUI Node

A ComfyUI custom node that runs the audio-separator command-line tool to split an audio file into stems (vocals, instrumental, drums, …).

Design: a GUI around the audio-separator CLI

This node wraps audio-separator and is visually designed around its command-line arguments. Every input maps 1:1 onto a CLI flag, and every button mirrors an operation you could run from a terminal — so if you already know the CLI, the node should feel familiar.

| Node field / button | audio-separator equivalent | | --- | --- | | input_files | positional input-file argument (e.g. /tmp/song.wav) | | output_dir | --output_dir | | model_filename | --model_filename | | command_options | any extra CLI flags, verbatim (see audio-separator --help) | | Show Command Help button | audio-separator --help | | List Loaded Models button | list the model files present in the models folder | | List Model Stems button | show a model's stems, numbered to match output_stempath_N | | Remove Models button | delete everything in the models folder (failsafe) | | STOP PROCESS button | kill the running audio-separator subprocess |

Under the hood the node builds and runs exactly this:

audio-separator --output_dir <output_dir> \
                --model_file_dir <models dir> \
                --model_filename <model_filename> \
                [your command_options…] \
                --custom_output_names '{"Vocals": "song-model-(Vocals)", …}' \
                --output_format wav \
                <input_files>

--custom_output_names and --output_format are added automatically only when you haven't supplied them in command_options, so your values always win.

Inputs

| Field | Type | Default | Notes | | --- | --- | --- | --- | | input_files | STRING | /tmp/example.wav | Full path to the audio file to separate. | | output_dir | STRING | /tmp/ | Directory where stems are written. | | model_filename | dropdown | — | A model from the audio-separator catalog (or one already in the models folder). | | command_options | STRING | (empty) | Extra CLI flags verbatim, e.g. --enable_float32. May include --custom_output_names '{"Vocals": "my_v"}' and/or --output_format flac to override auto naming. |

Outputs

| Output | Meaning | | --- | --- | | STDOUT | The CLI's console output for the run (also streamed live to ComfyUI's server log). | | output_stempath_1 … 8 | Full path of each stem file, in model order. Unused slots return an empty string. |

Chaining: each output_stempath_N is just a file path — feed it as the input_files of another Audio Separator CLI node to run a second separation on that stem. Chain nodes together to build complex multi-stage graphs (e.g. isolate vocals → then split those vocals into lead/backing).

Output naming

By default each stem is written as:

{input_stem}-{model_file_stem}-{(stem)}.{format}

e.g. input song.wav with model htdemucs_ft.yaml → song-htdemucs_ft-(Vocals).wav, song-htdemucs_ft-(Instrumental).wav. The number shown by List Model Stems (1, 2, …) is the same index as output_stempath_N. Override with --custom_output_names / --output_format in command_options.

Buttons

• Show Command Help — runs audio-separator --help and shows it.
• List Loaded Models — lists the model files currently present in the models folder.
• List Model Stems — resolves the selected model's stems (fetching only its small config, never its weights) and lists them numbered to match output_stempath_N.
• Remove Models — destructive: deletes every file in the models folder so unwanted or damaged models are purged and a re-download starts clean. A confirmation dialog guards it.
• STOP PROCESS — hard-stops (SIGKILL) any running audio-separator subprocess, e.g. one that has hung mid-download or mid-inference.

Installation & setup

1. Install ComfyUI.

2. Put this folder in ComfyUI's custom_nodes/ — or install it via ComfyUI-Manager, which reads the node's requirements.txt and installs audio-separator (the CPU baseline) into ComfyUI's environment for you.

3. If installing manually, run install.py with the same Python that runs ComfyUI so the dependency lands where the node looks for it (next to ComfyUI's interpreter). Pick ONE profile:

   | Profile | Installs | Command | | --- | --- | --- | | CPU (default) | requirements.txt — works on any machine, no CUDA needed | <comfyui-venv>/bin/python install.py | | GPU (CUDA ONNX) | requirements-gpu.txt — accelerates ONNX stems on a GPU | <comfyui-venv>/bin/python install.py --gpu |

   On Windows use the venv's ...\Scripts\python.exe instead of .../bin/python. You can also skip install.py and run pip install -r requirements.txt (or -r requirements-gpu.txt) directly. Do not combine both profiles — that leaves two ONNX runtimes and can break the CUDA provider. To point at a specific binary instead, set AUDIO_SEPARATOR_BIN.

4. Restart ComfyUI and add an Audio Separator CLI node (category audio).

Dependency notes: audio-separator is a heavy ML package (pulls PyTorch + ~28 others). The node was tested against 0.47.x; on any other version it logs a startup warning — run smoke_test.py before relying on stems. PyTorch-based stems already use your existing CUDA torch, so the CPU profile does not block GPU inference — only ONNX stems need the GPU runtime. audio-separator pins beartype<0.19.0; if other packages need a newer one, run pip install --upgrade beartype afterwards (pip warns about the conflict, but it works).

Models folder

Model weights/configs are stored in the node's models/ directory by default. Override its location with the AUDIO_SEPARATOR_MODEL_DIR environment variable. Models download on first use; Remove Models clears them all.

Notes & safety

• The CLI runs as a separate process, so a hung or crashed model can't take down ComfyUI — use STOP PROCESS to kill it and Remove Models to purge bad files.
• command_options is passed through verbatim; see the Show Command Help button for every flag audio-separator supports.