ComfyUI Karaoke

ComfyUI custom node pack for building karaoke pipelines. Currently covers the source and stem separation stages — fetch a track from the web, split it into vocal / instrumental stems, and hand the result off to downstream nodes. Future stages (pitch shift, lyric transcription, lyric sync, video render) will be added incrementally.

Ships four nodes:

• Video Downloader (yt-dlp) — fetch a URL to disk, output the file path. Supports cookie-based auth for Premium / members-only content.
• Load Audio (from path) — read a STRING path into ComfyUI's AUDIO type, so audio_only downloads compose directly with audio-consuming nodes.
• String → AudioPath — retype a STRING as AUDIOPATH for packs (e.g. UVR5) whose inputs demand that nominal type.
• Audio Separator — run MDX / VR / Demucs / MDXC models via python-audio-separator to split an AUDIO into two stems.

Features

• Download video (merged to mp4) or audio only (mp3/m4a/opus/wav/flac/aac/vorbis)
• Optional cookies.txt for Premium / age-gated / members-only content
• Path-based outputs that compose with VideoHelperSuite (VHS_LoadVideo) for video, and with the bundled Load Audio (from path) node for audio-only flows
• Stem separation with any model supported by audio-separator — models are auto-downloaded to ComfyUI/models/audio_separator/ on first use and cached across runs

Install

Drop this repo into ComfyUI/custom_nodes/ and run pip install -r requirements.txt in that directory.

For development, uv sync works instead.

CPU-only environments

requirements.txt pulls audio-separator[gpu] by default (CUDA). On machines without a GPU, install the CPU variant instead:

pip install audio-separator[cpu]

External dependency

audio_only mode requires ffmpeg on your PATH. Video merging to mp4 also uses ffmpeg. Install via your package manager (winget install Gyan.FFmpeg, brew install ffmpeg, apt install ffmpeg, etc.).

Usage

Video Downloader (yt-dlp)

Category: video/download.

Inputs

| Input | Type | Notes | |---|---|---| | url | STRING | The video URL | | mode | video | audio_only | Pick output kind | | output_dir | STRING | Defaults to ComfyUI/output/downloads | | filename_template | STRING | yt-dlp template, default %(title)s.%(ext)s | | audio_format | dropdown | Only used when mode=audio_only | | format_override | STRING | Raw yt-dlp -f string; overrides mode default when set | | cookies_file | STRING | Path to a cookies.txt (Netscape format) |

Outputs

• file_path (STRING) — absolute path to the downloaded file
• title (STRING)
• duration (FLOAT) — seconds

Load Audio (from path)

Category: audio. Reads an audio file at a given path and returns ComfyUI's standard AUDIO type — a dict of {"waveform": Tensor[1, channels, time], "sample_rate": int}. Sample rate is preserved as-is; no resampling.

• Input: audio_path (STRING)
• Output: AUDIO

Uses torchaudio.load() under the hood, with a soundfile fallback.

String → AudioPath

Category: audio/utils. Passes a STRING through unchanged but retypes it as AUDIOPATH, which is what some audio packs (notably UVR5 forks) declare on their inputs. ComfyUI's type matching is strict/nominal, so without this step a plain STRING won't wire.

• Input: path (STRING)
• Output: AUDIOPATH

Audio Separator

Category: karaoke/separation. Runs a source-separation model on an incoming AUDIO and returns two stems as AUDIO.

Inputs

| Input | Type | Notes | |---|---|---| | audio | AUDIO | Standard ComfyUI audio dict | | model_filename | dropdown | Model to run (pulled from audio-separator's manifest, falls back to a curated list) | | output_format | wav | flac | Intermediate file format used during separation |

Outputs

• primary_stem (AUDIO) — first stem the model produces (e.g. Vocals for a vocal model)
• secondary_stem (AUDIO) — second stem (e.g. Instrumental)
• model (STRING) — the selected model_filename, passed through so downstream save nodes can compose filenames like vocals_<model>.wav

Stem semantics depend on the model; the names are deliberately generic. Consult the model's documentation to know which stem is which.

Models are downloaded on demand to ComfyUI/models/audio_separator/ and reused thereafter. Outside a ComfyUI environment (e.g. running the smoke test standalone), the default cache location is ~/.audio-separator/models/.

Using a shared models folder (extra_model_paths.yaml)

If you already have a models folder mounted elsewhere (shared between UIs, on a separate drive, etc.), point ComfyUI at it via extra_model_paths.yaml:

my-setup:
  base_path: /mnt/models
  audio_separator: audio_separator/

The node searches every registered audio_separator folder for the selected model. If the file is already present in any of them, that folder is used directly (no re-download). If it isn't present anywhere, the first registered folder is used for the fresh download — entries in extra_model_paths.yaml load before custom nodes, so a configured external path takes precedence over the default ComfyUI/models/audio_separator/.

Typical wiring

Video Downloader (mode=audio_only)  ──►  Load Audio (from path) ──► Audio Separator ──► primary_stem (AUDIO)
                                                                                     └─► secondary_stem (AUDIO)

Video Downloader (mode=video)       ──►  VHS_LoadVideo ──► IMAGE + AUDIO
Video Downloader (mode=audio_only)  ──►  String → AudioPath    ──► UVR5 (external pack)

Using Premium / authenticated downloads

Export your browser cookies to a Netscape-format cookies.txt:

1. Install a cookie exporter extension (e.g. "Get cookies.txt LOCALLY" for Chrome/Firefox).
2. While logged in to the target site, export cookies for that domain.
3. Save the file somewhere ComfyUI can read (e.g. ComfyUI/user/cookies.txt).
4. Set the node's cookies_file input to that absolute path.

Security note: the cookie file grants session access to your account. Keep it out of shared folders and out of version control.

Caching behavior

ComfyUI caches node outputs by input hash. Same URL + same options = no re-download on repeated queue runs. Change the URL or delete the file manually to force a fresh download. The same applies to Audio Separator — identical waveform + model = cached stems.

Roadmap

Stages beyond stem separation are planned but not yet implemented: pitch shifting, lyric transcription (Whisper), lyric sync (LRC/SRT), and karaoke video rendering with burned-in lyrics.