AI Audio Artifact Reduction (experimental)

You generated a song with MiniMax Music 3 or YuE2, and there it is again: a thin whistle riding the vocal in the second chorus, or a metallic ping that isn't in any instrument you asked for. AI Audio Artifact Reduction turns those spikes down a few dB and hands you the removed audio so you can hear what it cost.

Two caveats up front. It's experimental, and it is not an AI-origin detector. It finds spectral outliers - brief, narrow things that stick out from their neighbourhood in both frequency and time. A soprano note swelling out of a quiet mix looks exactly like the bug you're hunting.

How it works

No model, no GPU - just a short-time Fourier transform. The audio is windowed at roughly 43 ms with 75% overlap, and every time-frequency bin gets scored against two baselines: a local frequency median about 350 Hz wide, and a temporal median about 300 ms long. A bin only counts if it beats both. That two-sided test is the trick: a sustained tone was already there a moment ago, so it fails the temporal test, and broadband noise has equally loud neighbours, so it fails the frequency test.

Sensitivity sets the bar: Gentle wants 12 dB of frequency prominence and 10 dB of temporal novelty, Balanced 9/8, Strong 6/6. Candidates get smoothed into a soft-edged mask, clamped to your band and your dB ceiling, and then the node reconstructs only the removed component and subtracts it from the source. No hard gating, no resynthesis, no length change, and all channels share one mask built from combined channel power, so anti-phase stereo can't hide from the detector.

The inputs you'll actually touch

Nine fields in the schema, but only a few are worth thinking about.

• mode - start at Analyze only. It reports candidates, passes audio through untouched, and outputs silence on removed_audio. Look at what it found before you let it cut anything.
• sensitivity - Balanced by default, and the right place to live. Drop to Gentle if you hear it eating your cymbals; Strong finds more, including things that were music.
• max_reduction_db - the ceiling, 3 dB default, 8 max. Small numbers are the point. Needing 8 means you want a spectral editor.
• mix - wet/dry. 1 is full reduction, 0 returns the original.

The rest: min_frequency_hz defaults to 3000, which protects bass and the body of the vocal. max_frequency_hz defaults to 18000 but is capped at 45% of the actual sample rate, so on a 32 kHz file your ceiling is lower than the number you typed - and it must exceed min_frequency_hz or the node raises. protect_transients backs the mask off around broadband attacks to keep percussion intact; enabled bypasses everything without touching your wiring.

Four outputs: audio, removed_audio, artifact_reduction_json, and a one-line info string. Wire removed_audio into a native PreviewAudio and listen to it alone - that's exactly the input minus what you got back, mix included, and the fastest way to catch the node eating a violin.

Where it goes

In the bundled YuE2 / MiniMax workflow it sits in a CLEAN group between optional Refinement and Mastering: generated audio → Refinement → artifact reduction → Auto-EQ → manual EQ → rate → dynamics → exports. It's toggled by artifact_reduction_enabled in CHOOSE, starts on at Balanced, and runs independently of Refinement and Mastering. In the Audio Enhancement Lab and classic MiniMax graphs, add it manually after restoration and before Auto-EQ.

Install

Through ComfyUI Manager, search MiniMax Music Production Toolkit, or:

cd ComfyUI/custom_nodes
git clone https://github.com/jplenio/ComfyUI-MiniMax-Music-Production-Toolkit
cd ComfyUI-MiniMax-Music-Production-Toolkit
python -m pip install -r requirements.txt

Use the Python that runs ComfyUI - on Portable that's the embedded one; the repo ships install_requirements.bat to find a nearby venv. Restart ComfyUI, then hard-refresh the browser.

The good news: this node needs no model downloads and no extra packages. Its dependencies are numpy, scipy, soundfile, imageio-ffmpeg, mutagen and Pillow - if the toolkit loads at all, artifact reduction works. The heavy stuff people complain about (the GGUF LLM, llama-cpp-python, FlashSR, SheetSage2) belongs to other nodes in the pack.

Where people get burned

False positives. Bright percussion, sibilance and legitimately piercing high notes all trip the detector. If something sounds duller afterwards, lower mix or max_reduction_db, raise min_frequency_hz, or go Gentle - and compare enabled versus bypassed at matched loudness before deciding.

It won't fix what it looks like it should. Wrong lyrics, a broken arrangement, whole distorted instruments, clipping, persistent hiss and broadly watery artefacts are outside its remit. Clipped peaks have their own node in the toolkit; for a consistently harsh top end, the YuE2 - Smooth highs EQ preset does more than this ever will.

Nodes missing after install. If the pack doesn't show up, the console says IMPORT FAILED - usually requirements.txt installed into the wrong Python. Restart fully, not just refresh.

Treat it as a scalpel for one defect class: run it after the model stops improving, before you shape the tone, and keep removed_audio patched in so you know what you gave up.