Stems (separate and mix)
Stems splits a rendered (or uploaded) song into four stems - vocals, drums, bass, other - and mixes them back together with your own balance: Separate Stems and the Stem Mixer.
song templates: … Render ─► Plenio · Master [ Separate Stems ─► Stem Mixer ─► Audio HQ ─► EQ ─► Loudness ] ─► Export
4 · Enhance & Master: Load Audio ─► Stems (optional) ─► Audio HQ ─► EQ ─► Loudness ─► Export| Template | Stems | To use them |
|---|---|---|
| 1 · YuE2 · Song, 2 · YuE2 · Cover, 3 · MiniMax · Song, 5 · YuE2 · DAW | in the Plenio · Master block, switch stems off | switch stems on on the Master node, then open the block (the icon at the top right of the node) to set the mix |
| 4 · Enhance & Master | the block Stems (optional) under Load Audio, bypassed | select it, press Ctrl+B, then open it |
The stems come before the rest of the mastering on purpose: Audio HQ, the EQ and the loudness then hear your mix, not the raw render.
In the song templates (since 2026-10-24) the stems are the first stage of the Plenio · Master block: its stems switch (off) decides whether the song goes through them. Off, the song goes straight to Audio HQ - the separation does not run and its model is not loaded. On, Separate Stems splits the song and the Stem Mixer mixes it back; open the Master block to find the mixer with its strips and waveforms in its STEMS group (the breadcrumb at the top of the canvas takes you back). A bypassed stage cannot stand inside the block - the mixer has no audio input a bypass could pass on -, so three If/Else switches there choose the mix or the song and the two reports; leave them as they are and use the block's switch.
In 4 · Enhance & Master the block is one collapsed node, Stems (optional) (a subgraph of the loader, Separate Stems and the Stem Mixer), under Load Audio. To use it: select it, press Ctrl+B, then open it with the icon at the top right of the node.
The residual: why a neutral mix is exact
A separation model yields at most four stems; whatever it cannot place - room tone, reverb tails, separation errors - would simply vanish. Plenio computes it explicitly:
rest = input − (vocals + drums + bass + other)
rest is mixed as a fifth strip, not thrown away. That gives the block's core guarantee:
With neutral settings (every gain 0 dB, nothing muted or soloed, no compression, no sends) the Stem Mixer returns the input sample for sample - same rate, same length. Separation errors can only ever affect what you changed yourself.
The separation report names the stems and how much energy the residual holds, so you can see how much the model left behind.
Separate Stems
| Input | the song (before mastering) and a Load Audio Model of kind separation |
| Output | the stems plus the residual, and the separation report |
| Contract | the input's rate and length, stems in float32 |
- The model works at 44.1 kHz stereo. The node resamples to that rate, separates chunk by chunk with crossfades (the checkpoint's own chunk size), and resamples the stems back - so a 48 kHz render stays 48 kHz.
- A mono input is fed to the model twice (left = right); nothing else is guessed.
- ComfyUI caches this node: changing only the mixer does not re-run the separation.
Stem Mixer
The node has one widget: a strip per stem plus rest. Under Advanced you can also see and edit the raw plenio.stem_mix/1 JSON - it is the same value the widget writes, never a second one.
| Per strip | Range | Notes |
|---|---|---|
| gain (fader) | −60 … +12 dB | −60 dB shows as -∞ |
| M / S | on/off | solo beats mute; any solo silences every other strip, including rest |
| compression | 0 … 1 | one knob: threshold −12 … −30 dB, ratio 1 … 6, 10 ms attack / 120 ms release, no make-up gain; the report names the gain reduction |
| reverb send | 0 … 1 | to the one shared reverb bus |
| delay send | 0 … 1 | to the one shared delay bus |
| muted ranges | [start_s, end_s] | drawn on the strip's waveform: drag to mute a stretch, click a range to remove it |
| save | on/off | saves this stem as its own 24-bit FLAC next to the song on the next run (default: off) |
Every strip is there before the first run - vocals, drums, bass, other and the residual rest - so you can set the balance for the run you are about to start. Dragging a fader writes the value live (the strip stays under your pointer) and the release updates the whole node; every control carries a tooltip.
A stem as its own file
A strip switched to save is saved on the next run, as a 24-bit FLAC named after the stem: Export Release writes it next to the song, into a folder named after the song with -stems - for output/plenio/My Song.flac the stems are output/plenio/My Song-stems/drums.flac, vocals.flac ... (a batch of takes: drums take 1.flac ...). The folder takes the song's number with it (My Song (2)-stems), so nothing is overwritten, and the release record lists every stem file. In the templates the Stem Mixer's report goes into the export; in a workflow of your own, connect it to one of Export Release's reports inputs - without that, the stems go to output/plenio/stems (drums.flac, then drums (2).flac). The file holds that strip's own signal: its gain, its compression and its muted ranges are applied; mute/solo and the shared buses belong to the mixdown, not to one stem.
Muted ranges cut a stem's content in that window with a 10 ms fade and are merged and clipped to the song. The song keeps its length - cutting time out of the song would misalign the stems, so it is not offered.
The bass and the vocals as MIDI
A saved bass or vocals strip is also written as a standard MIDI file next to its FLAC - bass.mid, vocals.mid (a batch: bass take 1.mid ...) - for a DAW (a user's request, October 2026: "more MIDI tracks - bass or other lead instruments"). Experimental.
- The notes come from the stem itself: its pitch (the YIN tracker of Sung Pitch, in the bass's or the voice's range) and its attacks - a note played again on the same pitch is a note of its own, a vibrato is not. The velocity follows the level.
- The notes stay where the stem plays them - no grid: a YuE2 render runs a little ahead of or behind its score, and the stems in the DAW are at the recording's times too. The file's tempo and meter are the song's (its score's; else the brief's tempo; else 120 BPM in 4/4), so in a DAW at that tempo every note sounds with its stem. Quantise there if you want the grid.
- Measured on five renders (pop, electronic, rock, hip-hop, singer-songwriter): 93-100 % of the bass notes' time lies on a tone of the score's chord there - the pitches are right; the rhythm is the stem's, including its looseness. About 2 s per stem for a 3-minute song.
- Only one voice at a time can be transcribed: the drums and the other stem (guitars, keys, pads - several notes at once) have no MIDI file. The score's own MIDI (Song Sheet → Export MIDI: Vocal, Instrument, Chords) is the plan the render followed; the stems' MIDI is what the render played.
- Without an Export Release taking the mixer's report, the files go to
output/plenio/stemsat 120 BPM (the notes keep their times all the same).
The effect buses
| Bus | Model |
|---|---|
| Reverb | convolution with a synthesised, seeded, decorrelated stereo impulse response; presets room (1.2 s), plate (1.6 s), hall (2.4 s) |
| Delay | stereo feedback delay; time in ms or a note value at 120 BPM; the first echo has the send level, feedback 0 ... 0.8 sets how much of each echo returns (0: a single echo); low-pass in the loop |
Both are deterministic: the same settings give the same result. A send without its bus is refused - the mixer stops with a message instead of silently ignoring your send. Raising a send in the widget writes the documented default bus settings along with it, so the first raise just works.
The mixer's waveform of the last run comes from the separation that produced the stems; before the first run the strips show the documented stems and an empty waveform, and the node says so.
Mixdown contract
- Float64 inside, no normalisation: the mixer never changes your loudness on its own.
- Output length = input length; effect tails are cut at the song end with a 50 ms fade.
- Peaks and per-strip gain reduction are reported; Plenio · Master sets the loudness afterwards.
- The mixdown and the separation each write a report; both reach Export in the templates.
Model and licence
| File | Folder | Size | Licence |
|---|---|---|---|
model_bs_roformer_ep_17_sdr_9.6568.ckpt | models/audio_separation | 527 MB | MIT (the trainer and the vendored inference code); the checkpoint was trained on MUSDB18-HQ, whose dataset terms are research-oriented |
The inference code is vendored in plenio/third_party/msst (pinned commit and per-file hashes in its NOTICE.md). The checkpoint carries its own config, so a second file is not needed. ComfyUI's missing-model dialog offers the file (its loader names the download); Plenio itself downloads nothing.
The 0.3 release ships BS-RoFormer as the only separation model. An alternative checkpoint with the same architecture can be dropped into models/audio_separation and chosen in Load Audio Model; other architectures are refused with a message rather than guessed at.
What is measured, what is not
- The authors report MUSDB18-HQ SDR 9.66 for this checkpoint. Plenio has not re-measured the separation quality itself.
- Everything the mixer guarantees (neutral = input, solo/mute rules, range merging, bus determinism, lengths and rates) is covered by tests that run without the model.
- The listening check (L2) on real songs with the real weights is still to be done; until it is accepted, the block is marked experimental.
tools/studies/stem_study.pyproduces the numbers (speed, VRAM, the neutral-mix and muted-range contracts) and - with--write-audio <dir>- the material for the verdict: the neutral mix, the four stems, the residual and one example mix as 24-bit FLAC. The checklist is inCURRENT_STATUS.md(L2) and the plan §6.
Limits
- Separation quality is the model's, not Plenio's: dense mixes with wide stereo vocals are the hard case, and "other" holds whatever did not fit the three named stems.
- A perfectly clean vocal is not possible this way. YuE2 renders one mixed track - there is no vocal track of its own to export - so the vocals are always separated, and a separation always leaves traces (measured, October 2026: separating the vocal stem a second time with the same model changes it by -46 dB, i.e. not at all; muting it between the sung lyrics lines would also cut ad-libs and backing vocals - what the vocal stem holds there is as pitched as the singing). What would help is a separation model made for vocals alone; that needs a new download and is not part of this release.
- The residual keeps a neutral mix exact - it does not repair a separation error once you change gains.
- Separate files are written only for the strips you switch to save; Export always takes the mixdown (and the reports). The saved file carries the strip's gain, compression and muted ranges, not the mute/solo decision and not the shared bus tails.
- Effect buses are shared: every strip's send reaches the same reverb and the same delay, which is what keeps the mix reproducible.