AI Audio Enhancer Pro

Your generated song comes out of ACE-Step sounding flat, thin, and a little lifeless. That's exactly the gap this node is built for. AI Audio Enhancer Pro is the "make it sound like a real record" pass you drop after a music or speech model - it separates the audio into stems, brightens, compresses, and glues it back together into something you'd actually post.

It sits squarely in the audio branch of ComfyUI's graph. Anything that emits an AUDIO - a TTS node, a music model like ACE-Step, or a LoadAudio node - feeds in, and you get the same AUDIO dict out, ready to wire into PreviewAudio or a save node. The author's own README frames it "Use With ACE Step," and that's the honest use case: AI music generators nail structure but routinely spit out clipping, muffled mids, and artifacts. This is the cleanup.

How it actually works

Two paths, chosen behind the scenes:

With source separation (the good path). It runs the audio through Demucs (Meta's stem-separation model, htdemucs by default), splitting the mix into vocals, drums, bass, and "other." Each stem gets its own enhancement pass - vocals get a presence boost around 3.5kHz plus a de-ess cut, drums get transient enhancement, bass gets warmth, and the rest get a blend - then everything is recombined and limited. That's why you get separate vocals_enhance, drums_enhance, bass_enhance, and other_enhance sliders: you can brighten the voice without touching the beat.

Without it (the fallback). If Demucs isn't installed, or you flip use_source_separation off, it falls back to plain frequency-band processing - a simple_mode of "Standard" or "Aggressive," with the same clarity/dynamics/warmth/air knobs doing multiband EQ and compression on the whole mix instead of per-stem. Faster, no model download, but it's a broad-stroke EQ rather than surgical.

enhancement_level is the master fader - every other effect gets multiplied by it, and at 0 it returns the audio untouched. The tonal knobs map to real audio concepts: clarity is a mid-band presence boost, warmth a low-shelf lift, air a high-shelf sparkle above 10kHz, dynamics compression and transient shaping. dolby_effect widens the stereo image. If apply_limiter is on (it is by default), a pedalboard limiter at −0.5dB catches anything that would have clipped; otherwise it peak-normalizes.

The inputs that matter

Most of the 15 inputs have sane defaults, so a beginner can just set enhancement_level and go. The ones worth touching:

• enhancement_level - master intensity, 0–1. Start at 0.5 and nudge up. More isn't better; past ~0.8 things start to sound overcooked.
• device - defaults to cuda. If you're CPU-only, switch this to cpu or every run will error out on model load.
• The stem sliders - vocals_enhance / drums_enhance / bass_enhance / other_enhance. Boost vocals for podcast dialogue; boost drums and bass for a club track.
• use_source_separation - leave on if Demucs is installed; it's the whole point.

The single output, audio, feeds straight into PreviewAudio or downstream nodes.

Installing it

Grab it through ComfyUI Manager (search "ComfyUI-Audio_Quality_Enhancer") or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/ShmuelRonen/ComfyUI-Audio_Quality_Enhancer
cd ComfyUI-Audio_Quality_Enhancer
pip install -r requirements.txt

The pack's requirements.txt lists torch, numpy, soundfile, scipy, and librosa - then demucs and pedalboard as "recommended." Read that as "required for this node to do its job." Without demucs the model dropdown literally becomes "Not Available - Install Demucs," and without pedalboard the limiter and EQ passes silently degrade. Install them both:

pip install demucs pedalboard

Demucs downloads its weights (~80–300MB depending on model) on first use into the torch cache, so the first run is slow. You don't need SoX for this node - that's the other node in the pack.

Common issues

• It returns audio but sounds identical. Either enhancement_level is ~0, or - the sneaky one - Demucs isn't installed and it's silently running the fallback. Check the console: the code prints which path it took.
• CUDA errors on startup of the node. device defaults to cuda; flip it to cpu if you don't have a working GPU.
• Very short clips don't get separated. The code only attempts Demucs on audio longer than 3 seconds; anything shorter goes straight to fallback processing.

One honest caveat: this pack has almost no community footprint yet - a handful of threads mention it inside ACE-Step workflow shares, nothing like the install-war-story piles around bigger packs. It's a small, single-author tool that does one thing. Keep your expectations at "cleaner, punchier mix," not "studio mastering," and it delivers exactly that.