Audio HQ
Audio HQ is three refinement stages that run on every take before the EQ, inside the Plenio · Master block:
render ─► Plenio · Master [ Stems (switch: off) ─► HQ 1 ─► HQ 2 ─► HQ 3 (bypassed) ─► EQ ─► Loudness ] ─► Export| Stage | What it does | In the templates |
|---|---|---|
| HQ 1 · Stereo Cleanup | cleans the side signal (the stereo difference) above 2 kHz - the unsteady shimmer AI renders leave at the edge of the stereo image - and adds a little harmonic warmth; the middle (voice, kick, snare, bass) is not filtered | on |
| HQ 2 · Regenerate Highs | erase and regenerate: the UniverSR model (229 MB) generates new highs from the song without its top octaves; above 8.2 kHz - in every preset - they replace most of the original's, at the original's level | with HQ mastering on (off by default; on in 4 · Enhance & Master) |
| HQ 3 · De-Reverb | takes out the short metallic room after syllables and hits and smooths the phase above 2 kHz - but it also lowers the highs by 3-4 dB | bypassed (experimental) |
All three keep the song's sample rate and length and normalise nothing; Loudness & Dynamics after the EQ sets the level and the peaks. Every stage writes a report into the release record (profile, why it was chosen, settings, what changed) and a one-line summary on the node.
HQ mastering: the switch at the top
UniverSR is what makes HQ 2 sound best - and what takes the time: it about doubles a song's time (measured on 1 · YuE2 · Song, 97 s of music, RTX 5060 Ti: 280 s with it, about 130 s without). So every template has the HQ Mastering card at the top of its first group (beside it the Cover Art card), off by default in the song templates and on in 4 · Enhance & Master (since 2026-10-24) - click it to switch:
- on (on: less AI shimmer, 2× time) - the mastering at its highest quality: HQ 2 regenerates the highs. It can clearly reduce AI shimmer and artifacts. The UniverSR file (229 MB,
models/audio_sr) is needed. - off (off: faster, no UniverSR) - HQ 1, the EQ and the loudness run as always; HQ 2 passes the song unchanged and says off in its badge. UniverSR is not loaded and the file is not needed.
The switch is wired into the Master block only (in 4 · Enhance & Master into HQ 2), so switching it after a run runs only the mastering again - the song stays. In App mode it is one of the first choices.
To change a stage, open the Master block (the icon at the top right of the node), then choose a preset, or select a stage and press Ctrl+B to bypass it or switch it on. In 4 · Enhance & Master the three stages are plain nodes on the canvas (there is no Master block), and their presets are also in App mode.
Presets and auto
Every stage has the same five choices, plus custom for its own fields:
| Preset | For | In short |
|---|---|---|
| auto - from the song's genre | every template with a brief (the default) | picks one of the four below from the brief's genre |
| standard - pop, hip-hop, soul, world | the default sound | the proposal's values, as in the listening test |
| gentle - classical, jazz, folk | natural rooms and delicate highs | a little less side cleanup and warmth, 80 % new highs, a lighter de-reverb |
| strong - rock, metal, punk | dense, distorted highs | more side cleanup and warmth, 95 % new highs, a stronger de-reverb |
| wide - electronic, dance, ambient | intended stereo width | keeps more of the side, the new highs follow the original's level more closely |
HQ 2 always regenerates from 8.2 kHz with UniverSR, the setting of the listening test: the presets change only how much of the new highs is used and how closely they follow the original's level. (Until 2026-10-22 gentle and wide started at 12.2 kHz; on the templates' default singer-songwriter song that changed too little to hear.)
How auto chooses: it reads the brief's genre for known words - rock, metal and punk words make it strong; electronic, techno, house, trance, drum & bass, synth, ambient and the like wide; classical, orchestral, piano, jazz, folk, blues, country and the like gentle - checking in that order, so folk metal is strong and synth-pop wide. Without such a word it takes the family of the brief's template (metal/... strong, edm/... wide, classical/... gentle, ...), and otherwise standard - so the templates' default song (pop/singer-songwriter) runs standard. Without a brief (Enhance & Master, or the node on its own) auto is standard. The summary always says which profile ran and why, for example strong (auto: genre 'thrash metal' (metal)).
The numbers of each profile are in the stages' help pages. Only standard has been listened to; the other profiles are reasoned starting points and are marked provisional in the summary until they have been listened to.
Time and memory
HQ 1 and HQ 3 take a few seconds per minute of music on the CPU. HQ 2 takes the time: UniverSR runs each channel through four flow steps (two network passes each, both twice for its guidance) per 5-second chunk - on an RTX 5060 Ti about 1.5 seconds per second of music, so a four-minute song takes about six minutes more (much longer without a GPU). It runs in bfloat16 (a quarter faster than float32, the same result within -25 dB) and needs about 3 GB of VRAM. Faster on purpose: HQ 2's model steps (an advanced field) - 3 steps about 1.1 s per second of music, 2 about 0.75 s - give other, somewhat quieter highs (measured: -15 dB and -6 dB from the four-step highs); fewer steps were not listened to. For quick drafts leave HQ mastering off; its seed is fixed, so with it on ComfyUI reuses HQ 2's result as long as nothing before it changed.
Where it comes from, and what was measured
The stages follow a proposal (Plenio-AudioHQ v3.0, 2026-10-21); Plenio's code reproduces its DSP (checked to -140 dB). Three things differ on purpose:
- HQ 2 uses UniverSR. The proposal named a FlashSR model; in the listening test UniverSR sounded better than FlashSR (2026-10-22), so Plenio uses UniverSR only (MIT code, CC BY 4.0 weights, part of Plenio since Refine) and offers no FlashSR.
- No soft clipper. The proposal ran
tanhover the whole song when a peak reached 0.999 - seven of nine YuE2 takes peak at 0 dBFS, so it would have saturated almost every song. Loudness & Dynamics handles the peaks after the EQ. - No silent fallback. Without its model the proposal quietly replaced HQ 2 by a DSP trick that cut the highs above 10 kHz by about 15 dB. In Plenio, HQ 2 with HQ mastering on needs the model; a missing file is an error with a download hint (off, the file is not needed).
Measured on 45-second excerpts of nine YuE2 takes (standard):
| HQ 1 | HQ 2 (UniverSR) | HQ 3 | |
|---|---|---|---|
| below 6 kHz | ±0.4 dB | unchanged | -1 to -3 dB above 1 kHz |
| 6-10 kHz | -0.4 dB | -1 to -3 dB | -3 to -4 dB |
| 10-16 kHz | -0.4 dB | -3 to -8 dB | -3 to -4 dB |
| side above 10.5 kHz | -10 to -13 dB (narrower) | -4 to -13 dB (UniverSR makes both channels alike) | ±1 dB |
| attacks | unchanged | 0 to -1.5 dB | -1.5 to -4 dB, like the tails |
Together, HQ 1 and HQ 2 make the top octave clearly narrower (side -8 to -19 dB above 10.5 kHz). The listening test's verdict: HQ 1 very good, artefacts well removed without hurting the highs; HQ 2 very good, with UniverSR better still; HQ 3 rather dull. The listening material and the numbers are in the test report.
Refine and Audio HQ
Refine (48 kHz), which until 2026-10-21 ran in 3 · MiniMax · Song and sat bypassed in the other templates, is in no template any more - all audio work after the render lives in the Master block now, and HQ 2 regenerates the highs. Refine stays in the node library for a plain 48 kHz conversion; see Refine (48 kHz).