de-artifact

01What de-artifact does

It is a restoration tool, not a tone shaper. It looks for the specific kinds of damage that generative models and codecs leave in a signal, and subtracts them.

Music that comes out of a generative model (Suno, Udio, Stable Audio and friends) or out of a stem separator carries a distinctive residue. It is not hiss and it is not distortion — it is a smeared, slightly metallic layer that clusters in the upper spectrum, plus ringing around transients and sustained tones. Low-bitrate MP3, AAC and Opus leave a related artefact class for the same underlying reason: both are lossy representations reconstructed from a compressed code.

de-artifact attacks that residue directly:

• HPSS analysis splits the signal into a harmonic stream (vocals, synths, sustained tones) and a percussive stream (drums, clicks, transients), so the two can be treated with different intensity.
• A learned UNet mask — the ArtifactNet model, embedded in the plug-in — estimates, per frequency bin, how much of the energy there is artefact rather than music.
• Adaptive spectral subtraction removes that estimate in the power domain, with attack/release smoothing so the reduction moves musically instead of chattering.
• Automatic HF air recovery measures how much high-frequency energy the process removed and restores it on the output, so cleaning does not cost you the top end.

Because artefacts concentrate above roughly 3 kHz — where natural music has comparatively little energy — the model is conservative on genuine acoustic material by construction. It will not "clean" a well-recorded acoustic guitar into mush.

When to reach for it

• Full mixes generated by Suno / Udio / Stable Audio.
• Stems pulled out of a mix by a separator (htdemucs, Spleeter and similar) — especially vocals.
• Sources you only have as a low-bitrate MP3/AAC/Opus file.
• A final polish pass on a master that contains any of the above.

02Install

macOS

Download de-artifact-0.3.6.pkg from the download page and double-click it. The package is signed and notarized by Apple, so Gatekeeper lets it through without a warning. Follow the prompts.

One package installs all three formats:

The installer also clears the quarantine flag and fixes executable permissions on the Audio Unit, so it validates cleanly the first time Logic scans it.

Windows

Download either de-artifact-windows-x86_64-0.3.6-setup.exe (installer) or the matching .zip for a manual install. Windows ships VST3 and CLAP; Audio Unit is a macOS-only format.

• C:\Program Files\Common Files\VST3\de-artifact.vst3
• C:\Program Files\Common Files\CLAP\de-artifact.clap

System requirements

03First run

In your host's plug-in browser, look under:

• VST3 / CLAP — category Restoration, vendor Intrect, name de-artifact.
• Audio Unit (Logic / GarageBand) — Audio Units → Intrect → de-artifact.

Insert it on a track, a bus or the master. It processes in stereo. The plug-in reports its latency to the host, and the host's plug-in delay compensation lines everything back up automatically — see §10 for the exact figures.

On the first launch with an internet connection the plug-in requests a 14-day trial automatically; nothing is gated at install time. §11 covers the details.

04The interface

One window: a header strip, a large analysis display, and a control sidebar. Drag the bottom-right corner to resize — the aspect ratio is fixed, so the layout scales rather than reflows.

Every screenshot on this page is the plug-in processing a real AI-generated track — the curves are its actual output on that material, not an illustration.

4.1 Header

4.2 The artifact-reduction display

The vertical axis is reduction in dB, not signal level: a flat line on zero means "not removing anything here", and the further the curve dips, the harder the model is cutting that band. The horizontal axis is logarithmic frequency.

Emphasis handles (advanced)

Four draggable dots sit on the zero line. They form a hidden four-band EQ (±12 dB) that biases where the model concentrates — pull a band up to make it work harder there, down to make it leave that region alone.

• Drag horizontally — move the band's centre frequency.
• Drag vertically — set the emphasis gain (±12 dB).
• Right-click — reset that band.

4.3 Control sidebar

Every knob and slider behaves the same way:

• Drag up/down to change the value.
• Hold Shift while dragging for fine adjustment.
• Double-click to reset to the default.
• Hover for an inline description.

05Quick start

Five steps. Everything after this section is refinement.

1. Insert de-artifact on the track you want to clean.
2. Pick a preset from the header. Default is the right starting point for almost anything.
3. Play the material and watch the yellow reduction curve — that is what is being removed, and where.
4. Switch on Residual and listen. You should hear hash, fizz and ringing. If you can clearly hear the vocal performance or drum hits in there, you are over-processing — lower Strength.
5. A/B against Bypass to confirm the change is actually an improvement before you commit.

06Presets

Eight factory presets, plus any you save yourself. Click the preset name in the header to open the browser.

Factory presets

Picking one by source

Suno / Udio AI-generated full mix       →  Default, then Aggressive if needed
AI-separated vocal stem (htdemucs …)    →  Vocal
AI-separated drums / bass               →  Default or Aggressive
MP3 / AAC / Opus low-bitrate source     →  Conservative
Finished mix or master bus              →  Soft Master Bus
General mastering polish                →  Balanced
Not sure                                →  Default

Loading a preset sets every control at once. After that each control is free — nudging one does not snap you back to the preset, it just marks the preset as modified with an asterisk.

Saving your own

In the browser, use Save As…, give it a name, and it is written to disk. User presets appear under the factory ones and can be deleted from the same panel (with a confirmation). You can also Import and Export presets as .json to move them between machines or share them.

• macOS — ~/Library/Audio/Presets/Intrect/de-artifact-rt/User/

07Controls reference

Every control, what it does, and what it costs you when pushed.

7.1 MONITOR

7.2 RENDER

7.3 TUNE — the everyday controls

Automatic HF air recovery (always on)

The model removes energy in the highs, because that is where artefacts live — which by itself would leave tracks sounding slightly dull. de-artifact measures how much high-frequency energy the process actually took out (dry versus wet, envelope-tracked) and restores it on the output automatically. You do not have to do anything; the top end stays open by default. Brightness only adds to or subtracts from that automatic restoration.

7.4 ADVANCED

08Reading the display

The yellow reduction curve is your main feedback loop. It is computed from the real output — dry against clean — at GUI rate, so it moves with the music and responds immediately to Strength, Detail, Attack and Release.

• Flat on zero — nothing is being removed at those frequencies. If the whole curve is flat while material plays, either Strength is at 0 or the model does not consider this signal artefacty.
• A deep dip in the top octaves — the normal, healthy shape for AI and codec material.
• Deep reduction low down — unusual. Check the residual: if you can hear bass or vocal fundamentals in it, reduce Strength or lower the corresponding emphasis handle.
• MAX — the peak reduction in dB, the single number to watch when comparing presets.

Because the curve deliberately excludes Brightness and the emphasis EQ, it always answers "how much is the model removing" rather than "what does the output sound like" — those are different questions, and mixing them into one display is what made earlier versions confusing.

The blue residual curve, shown with the Residual monitor, answers the complementary question: what was removed and where it sat in the spectrum. Together they tell you whether the amount and the target are both right.

09Workflow recipes

9.1 AI-separated vocal stem

1. Insert de-artifact on the vocal stem and load the Vocal preset.
2. Play, then switch to Residual. If breaths and consonants are disappearing into the residual, lower Strength.
3. For ringing sibilance, raise Resonance Depth to around 0.4 and push Harmonic toward 2.5.
4. If the result sounds dark, nudge Brightness up slightly — the automatic recovery has already restored most of it.

9.2 Suno / Udio full mix

1. Insert on the track or master and load Default; switch to Aggressive if the artefacts are heavy.
2. Use the MAX read-out to gauge how hard it is working, and keep checking the residual for musical bleed.
3. Raise Detail toward 40–60 % if you only want the most obvious artefacts touched and the rest left alone.
4. On dense mixes, try Separation: Strong so the Harmonic and Percussive controls target their material more precisely.

9.3 Lossy-codec residue (MP3 / AAC / Opus)

1. Insert on the master bus and load Conservative.
2. Switch to Residual — you should hear high-frequency hash and nothing else.
3. If the cleanup feels too subtle, step up to Balanced rather than raising Strength blindly.

9.4 AI-separated drums

1. Insert on the drum stem, load Default or Aggressive.
2. If transients start to soften, raise Transient protection.
3. For hi-hat ring and snare crackle specifically, raise Percussive rather than the global Strength.

9.5 Final master polish

1. Load Soft Master Bus on the master.
2. Keep MAX modest — a couple of dB is plenty at this stage.
3. A/B against Bypass at matched loudness before committing. On a finished master, restraint beats reach.

10Latency & performance

de-artifact runs its neural pipeline on a worker thread and reports the resulting latency to the host, which compensates automatically. The reported figure depends on the platform, because the macOS build primes one chunk less:

• Your DAW's plug-in delay compensation lines tracks back up. A short fill window at the very start of playback is normal and expected.
• Other sample rates are supported by internal resampling to and from the model rate. The latency in milliseconds stays the same, because the chunk and hop sizes scale with the host rate.
• The plug-in reports its exact value to the host — you never have to enter an offset by hand.

Apple Silicon acceleration

Since 0.3.4, the macOS build runs the UNet on the Apple Neural Engine through Core ML rather than on the CPU. On an M1 Max this cut UNet inference from about 65 ms to 4.4 ms per chunk and the whole pipeline from roughly 66 ms to 5.6 ms, taking real-time headroom from 2.7× to about 31×. Output was verified against the CPU backend by spectral null test at −63 to −71 dB — inaudible. If Core ML fails to initialise on a particular machine, the plug-in logs a warning and falls back to the CPU path rather than failing.

If you hear dropouts

1. Lower Strength — it is the cheapest change that reduces work.
2. Raise your host's audio buffer size (512 → 1024 or higher).
3. If, and only if, a bounce comes back with gaps, switch Offline Render on for that render and off again afterwards.
4. Freeze or bypass other heavy plug-ins in the chain.
5. On lower-power machines, real-time stereo is tighter — a larger host buffer is almost always the simplest fix.

11Licence

A 14-day free trial, then a paid licence. Nothing is gated at install time.

11.1 First launch

On the first launch with internet access, the plug-in requests a 14-day trial. The trial is bound to a one-way fingerprint of your machine and delivered as a signed licence file on disk — after that no network connection is needed for the rest of the trial. If the server cannot be reached on that first launch, a local fallback trial is issued so audio is never blocked, and the plug-in retries quietly in the background next time you open the editor.

11.2 The licence panel

From the panel you can Paste a key from the clipboard (useful in hosts that swallow keystrokes aimed at plug-in windows), Activate a key in the form INTRECT-XXXX-YYYY-ZZZZ, Sign out to free the seat before moving to another computer, and Buy to open the purchase page.

11.3 Files and privacy

• macOS — ~/Library/Application Support/Intrect/de-artifact/license.blob
• Windows — %APPDATA%\Intrect\de-artifact\license.blob

The file is a small signed binary readable only by your user account. The server stores a non-reversible machine fingerprint, the plug-in name, and trial/activation timestamps. No audio, no telemetry, no usage statistics. The plug-in contacts the server only for explicit licence events: first-launch trial, activation, refresh and sign-out.

11.4 Common licence questions

How do I move to another machine?

Sign out on the old machine, then paste and activate your key on the new one.

I lost my key.

Email [email protected] from the address you used at purchase and we will re-issue it.

Can I work offline?

Yes. Once activated, a paid licence runs offline until its refresh window comes round, at which point it needs one online check.

How many machines does one licence cover?

Three. If you hit the limit, the panel offers to release one of the existing seats.

12Troubleshooting & FAQ

I installed a new version but still see the old interface or behaviour.

The host cached the previous binary. Remove the plug-in from the track and add it again, or restart the DAW. Check the version number in the header to confirm which build you are actually running.

The processed audio sounds dull.

The automatic HF air recovery already restores most of the removed highs. If you want more, raise Brightness. You can also lower Strength, or pull the highest emphasis handle down so the model treats the top end more gently.

The first fraction of a second is silent.

That is the reported latency window (see §10). Your DAW's delay compensation handles it; it is normal.

Playback stutters periodically.

Check that Offline Render is off — it is meant only for bounces and makes the audio thread wait during live playback. Then raise the host buffer size. If it persists, send us the ⚙ → Copy diagnostic report output.

A bounce or freeze comes back with gaps.

Some hosts do not announce offline rendering. Turn Offline Render on, run the bounce, then turn it back off for normal playback.

Does it work in Logic Pro?

Yes. Install the macOS package and de-artifact appears under Audio Units → Intrect in Logic and GarageBand. The "no sound in Logic" and AU validation failures reported on earlier builds were fixed in 0.3.2 and 0.3.5 respectively — make sure you are on the current version.

Complete silence on an Apple Silicon Mac.

Fixed in 0.3.6. A Core ML initialisation failure could take down the plug-in's host process silently. Update to the current build, which falls back to the CPU path instead.

Pro Tools / AAX support?

Not supported. AAX requires Avid Developer Partner enrollment plus PACE/iLok, which does not fit the current activation model. Pro Tools users can run de-artifact in any VST3- or CLAP-capable host.

What is the difference between de-artifact and de-leak?

de-artifact removes codec and generation artefacts — the smeared, metallic residue inside the material itself. de-leak removes bleed: content from another stem or source that should not be in this track at all. Many AI-generated stems have both problems, and the two plug-ins are routinely used together.

Can I process files without a DAW?

The plug-in is the product. There is an offline command-line path in the underlying library for batch work, but it is a developer tool and is not part of the installer.

13Support

Open a ticket at intrect.io/support — that is the fastest route and it reaches us directly. If your problem involves glitches, dropouts or a host misbehaving, include the output of ⚙ → Copy diagnostic report together with your host name and sample rate; it usually tells us what went wrong on the first read.

• Platforms — macOS 11+ (Apple Silicon and Intel; VST3/CLAP/AU) and Windows 10+ (x86_64; VST3/CLAP). Linux is not packaged.
• Signing — the macOS package is signed and notarized; the Windows installer is not yet Authenticode-signed.