Audio Stem Split

Want the vocals off a track so you can remix it, karaoke it, or just isolate the drums for a reactive visual? Audio Stem Split is the one-button answer: feed it a song, get back vocals, accompaniment, drums, bass, and other - five separate audio outputs from one input. And here's the part that makes it special in this pack: the README promises it's "built in, nothing extra to install," and unlike most of the audio suite that's actually an AI model doing the work - a heavyweight one, at that.

The mechanism: it runs HDemucs - specifically torchaudio's bundled HDEMUCS_HIGH_MUSDB_PLUS pipeline, the hybrid-demucs model trained on the MUSDB set. Torchaudio ships the model class itself, which is why there's no extra pip install and no separate repo to clone. What it doesn't ship is the weights: on first Run, it downloads them into your torch hub cache. So your first stem split will pause for a download (hundreds of MB), and it will chew on CPU or GPU depending on where ComfyUI's torch lands - HDemucs is a real convnet and a full song split is genuinely compute-heavy, not instant like the FFmpeg-filter FX stages. After that first run, the weights are cached and it's just slow-in-a-normal-way.

Input is audio or video (it splits the video's soundtrack), and it has zero parameter widgets - this node is as close to "drop it in and Run" as ComfyUI gets. The five stems come out as vocals, accompaniment, drums, bass, other. Wire vocals into one Mix, the instrumental into another, or run drums through an Audio Reactive stage to drive visuals off the actual percussion.

Honest expectations: HDemucs at the high-MUSDB setting is very good, but "very good" still means the vocals output can carry ghostly bleed from backing instruments, and the other stem (guitars, keys, pads) is the muddiest of the five. If you're doing a professional remix you'll still reach for a dedicated GUI like Demucs or a paid tool; for a ComfyUI pipeline that needs a vocals-free bed or a drum track to trigger visuals, this is exactly right, and it's the cleanest stem split you'll get without leaving the canvas.

Install is the pack-wide story: ComfyUI Manager search "ComfyTV", or git clone https://github.com/jtydhr88/ComfyTV into custom_nodes/, then a full backend restart (Desktop/macOS: clone by absolute path into the running instance). ComfyTV declares zero extra pip dependencies - torchaudio comes with ComfyUI. Remember the first-run weight download; if the split errors at 0%, check the log for a hub download failure and your connection, not the node.