HQ 3 · De-Reverb

Here's a rare thing: a node that ships bypassed in every template it appears in, is flagged experimental in its own schema, and whose author's published listening verdict is "rather dull". HQ 3 · De-Reverb is the third and last Audio HQ stage, and you should go in knowing that. It's genuinely useful for one specific problem, and it's a trap if you switch it on expecting a room remover.

What it does

Two processes, applied per channel above 600 Hz.

De-reverb is a late-envelope subtraction: a delayed, smoothed copy of the spectrum stands in for the short metallic room that hangs after syllables and drum hits, and where that copy is strong relative to the direct sound, the spectrum gets turned down - up to 10 dB, at full strength from 3.1 kHz upward. Phase lock then pulls the phase of every spectral point above 2 kHz toward the magnitude-weighted phase of its neighbours, reaching at most 47% from 4.6 kHz up. Steady tones are left alone; smeared, incoherent ones get pulled into line.

Same sample rate and length out, and nothing normalised.

What it actually sounds like, measured

This is where the node gets interesting, and not entirely in a good way. On nine YuE2 takes at the standard profile it takes off 3 to 4 dB above 3.2 kHz, with attacks and tails lowered alike - so it behaves more like a gentle high cut than like a dereverberator. On a test song with no reverb at all it removed just as much. That's why it's bypassed in the templates and why the manual calls the result dull.

So: don't reach for it as your first move. Reach for it when a take genuinely has that short metallic boxiness around the vocal - the kind of tail that sounds like the singer is standing in a tiled bathroom two feet across - and you've already tried HQ 1 and the EQ. If it works, it really does. If the take was clean, you just dulled your song for nothing.

The inputs and outputs

audio is the song; every item of a batch goes through. preset is the same five-way choice you'll see on HQ 1 and HQ 2 - auto, standard, gentle, strong, wide, custom - with two numbers behind each:

• standard: de-reverb 45%, phase lock 55%
• gentle: 25% / 30%
• strong: 65% / 70%
• wide: 30% / 40%

auto picks by genre from the optional brief input (Song Brief or Cover Brief), falling back to standard when nothing's connected. custom exposes the two fields so you can sit between profiles. All profiles are provisional - the node's summary says so out loud, and there's no point pretending otherwise.

Outputs are audio, the drier song, and report: profile, settings, and the highs before and after. The report is worth wiring into Export Release - it lands in the .plenio.json release record, so six months from now you'll know which take had de-reverb on and why it sounded flat.

Installing it

You get this with the pack; there's no model file involved and no extra Python:

cd ComfyUI/custom_nodes
git clone https://github.com/jplenio/Plenio-Music-Production-System

Or install Plenio Music Production System through Manager → Custom Nodes Manager and restart. Then run template 0 · System Check to make sure the template prerequisites are in place. In the song templates HQ 3 sits inside the Plenio · Master block, already bypassed - open the block with the icon at the top right of the node, select the stage, and press Ctrl+B to enable it. It's CPU work, a few seconds per minute of music, so cost isn't the reason it's off. The author's ears are.

Where people get burned

The obvious one: turning it on and leaving it on, because the song sounds "smoother" afterwards. It doesn't sound smoother, it sounds quieter up top, and you only notice once you A/B. Bounce a version with and without before you commit.

The second one is expecting de-reverb to rescue a bad room. This is a short-tail tool on a full mix; it is not RX's dialogue de-reverb and it can't separate a reverb tail from a cymbal. If you need that, do it on the stem - Plenio ships stem separation for exactly this kind of surgical work, and a vocal stem through HQ 3 at gentle is a much fairer test than a whole song at standard.

And expect to be early. There is essentially no community discussion of this pack yet, so no thread will reassure you that what you're hearing is normal. Read the pack's own test report and the Audio HQ page in docs/user/concepts/ instead - they carry the measurements and the listening verdicts.