Why does the first ~1 second go silent on a freeze / bounce?

The plug-in reports about 350 ms at 44.1 kHz for host PDC, the same on macOS and Windows. During Live's freeze the host signals offline mid-stream, and the worker thread re-primes for one block. Output is correct from the first audible frame onward.

The Residual monitor sounds different on each play. Bug?

Phase of the residual depends on chunk-grid alignment — same SR + same transport start gives bit-identical phase. Spectrum and RMS are stable across runs; only the per-sample residual rotates.

The result sounds dull. What knob fixes that?

Raise Brightness. The model removes energy in the highs because that is where artefacts live, and de-artifact already measures how much it took out and restores it automatically. Above 0, Brightness adds neural bandwidth extension on top of that — a small network regenerates the 8–22 kHz band from the cleaned signal. Negative values darken instead, scaling the automatic recovery down. If it still sounds dull, lower Strength from its 60 % default, or load Gentle Reduction, which uses a soft mask floor for minimal colouration.

I'm getting dropouts on an M1 8 GB.

Lower Strength first — it is the cheapest change that reduces work — then raise your host's audio buffer (512 → 1024 or higher), which is almost always the simplest fix on a lower-power machine. Setting Separation back to Normal costs less CPU than Strong. Freeze or bypass other heavy plug-ins on the chain; total plug-in latency feeds back into the host's buffer math. Since 0.4 the build runs the artifact model frame by frame on the CPU on every platform, Apple Silicon included, and reports about 350 ms at 44.1 kHz. The 0.3.4-era Neural Engine chunk path — 2.7× to about 31× real-time headroom on an M1 Max — is still reachable, but it is no longer the default.

Will de-artifact lower ArtifactNet's AI probability?

No. de-artifact is designed to improve sound, not to minimize ArtifactNet or any other detector score. Processing changes the signal a detector receives, so the reported probability can move either up or down. A higher post-processing score does not mean the track became more AI-generated, and the score is not a measure of sound quality or how many audible artifacts were removed.

Is there a free trial?

Yes — 14 days of full quality with no gaps, every preset, every parameter. The same installer becomes the licensed build the moment you drop in a key. After the 14-day trial the plug-in inserts a brief silent gap every 60 seconds until you activate.

Do I need a plug-in to remove AI artifacts?

No. de-artifact also runs as de-artifact Cloud — full-length cleaning on our GPUs, driven from the browser — so you can remove AI artifacts from a full-length track without installing anything. The plug-in is the option for working inside your DAW with all 32 parameters under your hand.

AI Music Artifact Remover — de-artifact

What does de-artifact do?

de-artifact is a VST3 / AU / CLAP plug-in that removes the artifacts AI music generators leave in a full mix: RVQ ghosting, codec residue and high-frequency aliasing from Suno, Udio, Stable Audio, MusicGen and Riffusion. It splits the audio into harmonic and percussive parts and applies a learned mask to each, on the master bus, with a 14-day free trial.

Hear it on real tracks

Pick a track and tap Original / Cleaned / Diff to A/B in place. Playback is synchronized across all three — no re-buffer, no click, no time jump. Diff plays only what de-artifact removed.

0:00 0:00

Keyboard: 1 Original · 2 Cleaned · 3 or D Diff · M Mix/Vocal · Space play · ←/→ ±5 s

Tip: if Diff sounds like music, the preset is too aggressive — lower Strength in the plug-in. Mix is the full bounce; Vocal is the htdemucs-separated vocal stem.

New to de-artifact? Start with your own track.

Watch the 3-minute walkthrough: upload, process, compare Original and Cleaned, then choose browser cleanup or the 14-day DAW plug-in trial. The no-account listening preview is up to 30 seconds.

Hear a free 30-second preview on your own track

Is this the right tool?

de-artifact is a restoration tool, not a tone shaper. It looks for the specific damage that generative models and lossy codecs leave in a signal, and subtracts it. Everything below is what that does and does not cover.

Reach for it when

A full mix came out of Suno, Udio, Stable Audio, MusicGen or Riffusion.
You pulled stems out of a mix with htdemucs, Spleeter or similar — vocals especially.
The only copy you have is a low-bitrate MP3, AAC or Opus file.
You want a final polish pass on a master that contains any of the above.

Reach for something else when

The problem is bleed — a vocal audible in the drum stem, content that does not belong to that track at all. That is de-leak. The two are frequently used on the same source.
You are chasing one ringing resonance rather than a broadband residue. That is ENZYME.
You want to rebalance harmonic against percussive content deliberately. That is Intrect HPSS, which is free and macOS-only.
Your host is Pro Tools. There is no AAX build, so it will not load there.

What it will not do

Rescue a badly written track: It removes codec residue. Arrangement, performance and mix decisions are untouched, and a cleaned generation still sounds like the generation it came from.
Work hard on clean material — by design: Artifacts concentrate above roughly 3 kHz, where natural music has comparatively little energy, so the model is conservative on genuine acoustic sources by construction. On a bass stem it peaks at around 0.8 dB of reduction even at Strength 100 %. A near-flat reduction curve is a result, not a fault — reading it as a broken plug-in and cranking Strength is the most common way to damage a track.
Separate stems: It processes what you give it. Splitting a mix into parts is a separator's job, and cleaning the parts afterwards is what de-artifact is for.
Run everywhere: VST3, CLAP and Audio Unit on macOS 11+; VST3 and CLAP on Windows 10 2004 or later with an AVX2 + FMA CPU and the Microsoft Visual C++ 2015–2022 x64 runtime. No AAX, no standalone application, and no Linux build.

Try it free for 14 days, buy when you're sold

14-day full trial — every preset, every parameter, no gaps. After 14 days the plug-in inserts a brief silent gap every 60 seconds until you activate.

Start the 14-day trial

14-day free trial

Full quality, every preset, every parameter, no gaps for 14 days. After that the plug-in inserts a brief silent gap every 60 seconds until you activate with a license key.

de-artifact-0.4.3.pkg macOS Universal · 101 MB de-artifact-windows-x86_64-0.4.3-setup.exe Windows installer · 22 MB de-artifact-windows-x86_64-0.4.3.zip Windows portable · 44 MB

macOS — Apple Silicon + Intel · Windows — x86_64 with AVX2 · Linux — roadmap

Buy de-artifact

Release price — ends Dec 31

One-time purchase, three machines, free updates within v1.x. The license key arrives by email and unlocks the plug-in instantly — whether you're on day 1 of the trial or past day 14.

$179$149one-time

Try online →

Not sure yet whether your track was AI-generated? Run it through the free AI music detector — no account, no upload of the master.

Try it on your own audio

Drop any AI-generated audio and hear a 30-second A/B preview — free, no account needed. For full-length files there is de-artifact Cloud, or the plug-in itself.

WAV / MP3 / FLAC · max 50 MB · first run may take ~60s to warm up the GPU

🎵

Drop your audio file here

or click to browse

Processing

Submitting to GPU…

Preview ready

First 30 seconds · AAC 128k

Original

Three ways to run it

Same engine in all three. What changes is where it runs and how much of it you get to steer.

The plug-in

$149 one-time

In your DAW, on the 2-bus, with all 32 parameters under your hand. No per-minute cost and no connection required once it is activated.

Or start the 14-day trial →

Free online

Free · no account

Upload one track and hear the result. Good for deciding whether this is the right tool before you spend anything.

Try a track →

de-artifact Cloud

From $0.18 / minute

Full-length cleaning on our GPUs, from the browser. One tuned preset rather than the full control surface — that part is what the plug-in sells. Prepaid dollar credit, and Artist+ subscribers get $25 a month of credit included, shared with detection. An optional Clean + restore highs mode also generates the missing high end, at $0.75 per audio minute.

See rates →

14 factory presets

Every preset sets all controls at once; nudging a control afterwards marks the preset modified rather than snapping back. Save your own with Save As….

Preset	Use it for	Character
Conservative Universal	Subtle lossy-codec residue (MP3/AAC/Opus hash)	Barely touches the signal. Maximum transparency.
Balanced Universal	General polish with a gentle hand	Light, even cleanup.
Default default Universal	General AI-generated / AI-separated material	The starting point — Strength 60 %.
Aggressive Universal	Heavily artefact-laden AI tracks	Full strength plus resonance suppression.
Vocal Vocal	AI-separated vocal stems	Harmonic-weighted, transient- and sibilance-aware.
Gentle Reduction Universal	Cleanup where transparency matters most	Very light touch with a soft mask floor — minimal colouration.
More Harmonics Universal	Watery sustains and vocal smear	Maximum harmonic-channel reduction.
Soft Master Bus Master	Light artefact control on a finished mix	Wide, gentle cleanup — soft knee, slow release, strong resonance suppression.
Bring Your Breaks Bright Again Universal new in 0.4	Old break / loop / acapella samples whose top octave was cut off	Strong separation, maximum harmonic reduction, no spectral floor, Brightness +0.54.
Nicerizer Universal new in 0.4	AI-generated full mixes that need both the shimmer removed and the air put back	Firm reduction with Brightness +0.88 — the bandwidth-extension stage regenerating the top band.
Bmore Universal new in 0.4.3	Break / loop material that needs the separation without added top end	Strong separation, percussive channel closed, no spectral floor, no Brightness.
de-chirp Universal new in 0.4.3	Swept, tonal residue — chirp-like codec artefacts	Percussive-weighted reduction with a soft floor and Brightness +0.80.
Suno V6 Master new in 0.4.3	A full Suno V6 mix with generation residue across the whole track	Strongest of the set (Strength 1.60), full HF air, no spectral floor, 12 dB knee.
Stem Separation Stem new in 0.4.3	A stem you already separated, where heavy processing would show	Gentle (Strength 0.60), Strong separation, percussive channel closed, soft floor.

Features

Host-compensated latency

350 ms at 44.1 kHz (15,429 samples) — the same figure on macOS and Windows since 0.4.1, and the same at every sample rate in milliseconds. Reported to the host, so plug-in delay compensation lines the session back up.

4-band emphasis EQ

Drag spectrum handles to tell the model where to focus. The boost is fed only to the model — listener audio is auto-inverse-EQ'd, so tone is preserved.

Two monitor toggles

Bypass for A/B with the reported latency preserved, so the comparison stays sample-aligned. Residual plays only what is being subtracted — if you hear musical content in it, you are over-processing.

Dynamic resonance suppression

A soothe-style stage that tames narrow ringing peaks standing above the local spectral baseline. Depth, Selectivity and Threshold set how tight a peak must be and how far it has to rise before it counts — sibilance and harshness handled musically, rather than by a fixed band.

Separate harmonic and percussive control

The HPSS split gives the harmonic channel (vocals, synths, sustains) and the percussive channel (drums, transients) their own reduction strengths. Separation switches the median kernel between Normal (17) and Strong (31) for dense mixes.

How it works

de-artifact runs three stages in series. The audio path stays the model's domain until the very end — clean residual subtraction with no manual EQ tricks.

# Per-frame, real-time
audio (44.1k) → STFT → HPSS (median filter) ┬─ H_mag → ArtifactUNet → mask_H → soft_sub
                                       └─ P_mag → ArtifactUNet → mask_P → soft_sub
                                                                       ↓
                              clean = H_clean + P_clean → ISTFT → output

HPSS splits the spectrum into tonal and transient components. Each is masked independently by ArtifactUNet, a 3.6M-parameter forensic network trained on AI-generated music. The masks drive adaptive spectral subtraction — energy below the noise floor is removed, music above it is preserved.

System requirements

ArtifactUNet (3.6M params) runs on the audio thread alongside HPSS analysis, keeping a generous real-time margin even on entry-level Apple Silicon.

Requirement	Minimum	Recommended
macOS	Universal Binary — Apple Silicon (arm64) + Intel (x86_64) in one .pkg · macOS 11 (Big Sur) or later
Windows	Windows 10 2004 or later, 64-bit · CPU with AVX2 + FMA (Intel Haswell 2013+ / AMD Ryzen 2017+) · Microsoft Visual C++ 2015–2022 x64 runtime · .exe installer · VST3 + CLAP
CPU	Apple M1 / Intel Core i5 (4-core, 2.0 GHz+) / AMD Ryzen 5	Apple M2 or later / Intel Core i7 / AMD Ryzen 7
RAM	8 GB	16 GB+ (multi-instance comfortable)
Disk space	~150 MB (plug-in + ONNX models)	—
Audio buffer	≥ 256 samples	1024 samples for tracking
Sample rate	Any — rubato resampler 8 k–384 k. Internal model runs at 44.1 kHz.
Hosts (macOS)	VST3: Live, Reaper, Bitwig, Studio One · AU: Logic Pro, GarageBand, MainStage · CLAP: Reaper, Bitwig
Hosts (Windows)	VST3: Live, Reaper, Bitwig, Studio One, FL Studio · CLAP: Reaper, Bitwig
Logic / Pro Tools	AU supported, so Logic and GarageBand are covered. AAX (Pro Tools) is not supported — it requires Avid partner enrollment plus PACE/iLok, which does not fit the current activation model.
Linux	Roadmap.

FAQ

Why does the first ~1 second go silent on a freeze / bounce?: The plug-in reports about 350 ms at 44.1 kHz for host PDC, the same on macOS and Windows. During Live's freeze the host signals offline mid-stream, and the worker thread re-primes for one block. Output is correct from the first audible frame onward.
The Residual monitor sounds different on each play. Bug?: Phase of the residual depends on chunk-grid alignment — same SR + same transport start gives bit-identical phase. Spectrum and RMS are stable across runs; only the per-sample residual rotates.
The result sounds dull. What knob fixes that?: Raise Brightness. The model removes energy in the highs because that is where artefacts live, and de-artifact already measures how much it took out and restores it automatically. Above 0, Brightness adds neural bandwidth extension on top of that — a small network regenerates the 8–22 kHz band from the cleaned signal. Negative values darken instead, scaling the automatic recovery down. If it still sounds dull, lower Strength from its 60 % default, or load Gentle Reduction, which uses a soft mask floor for minimal colouration.
I'm getting dropouts on an M1 8 GB.: Lower Strength first — it is the cheapest change that reduces work — then raise your host's audio buffer (512 → 1024 or higher), which is almost always the simplest fix on a lower-power machine. Setting Separation back to Normal costs less CPU than Strong. Freeze or bypass other heavy plug-ins on the chain; total plug-in latency feeds back into the host's buffer math. Since 0.4 the build runs the artifact model frame by frame on the CPU on every platform, Apple Silicon included, and reports about 350 ms at 44.1 kHz. The 0.3.4-era Neural Engine chunk path — 2.7× to about 31× real-time headroom on an M1 Max — is still reachable, but it is no longer the default.
Will de-artifact lower ArtifactNet's AI probability?: No. de-artifact is designed to improve sound, not to minimize ArtifactNet or any other detector score. Processing changes the signal a detector receives, so the reported probability can move either up or down. A higher post-processing score does not mean the track became more AI-generated, and the score is not a measure of sound quality or how many audible artifacts were removed.
Is there a free trial?: Yes — 14 days of full quality with no gaps, every preset, every parameter. The same installer becomes the licensed build the moment you drop in a key. After the 14-day trial the plug-in inserts a brief silent gap every 60 seconds until you activate.
Do I need a plug-in to remove AI artifacts?: No. de-artifact also runs as de-artifact Cloud — full-length cleaning on our GPUs, driven from the browser — so you can remove AI artifacts from a full-length track without installing anything. The plug-in is the option for working inside your DAW with all 32 parameters under your hand.