ACENTAUR 1.2 — September 2026

Extract. A new mode that takes one instrument out of a Capture or a Generation and renders it on its own — vocals, drums, bass or any other track you pick. *Auto* runs the model's own extraction; *Custom* lets you describe the part and, if you like, draw a selection: only the selected passage goes to the model and only that passage comes back, in time with your arrangement.

A new panel. Settings and Advanced share one panel at the bottom left, the text fields carry their names as tabs, and TAX, Enhance Prompt and Generate stand together at the bottom right. The modes now run in the order you reach for them: Text to Music, Add Layer, Track Stack, Cover / Remix, Edit, Retake, Extract, Audio Analysis.

HP Filter. A zero-phase high-pass on every take, set in Advanced: 20, 40 or 60 Hz at 6, 12 or 18 dB per octave. Click a lit value again to switch it off. For vocals and guitars 60 Hz / 12 dB keeps the low end of your mix clear; for bass and drums leave it off or use 20 Hz / 6 dB. Every take also leaves ACENTAUR without the tiny DC offset the model used to put on it.

Use Capture. The button beside Clear hands the Capture you are working with to the model as the reference. In Cover / Remix it is on from the start — short Captures with the SFT and Base models come out far cleaner with it.

Easier lengths. The bar slider moves in single bars up to 32 and in larger steps beyond, so short loops are quick to set.

Clearer takes. The playback menu lists your Captures first and your Generations below them, and the arrows on either side of it step back and forth through your Generations with one click. Takes without a prompt are named after their mode, like *Add Layer 1*.

Analyze Audio. It can be cancelled, shows that it is working even right after a take, and finishes in a fraction of the time: the key is heard directly from the audio, the language from the lyrics.

Honest status line. When Generate rests, the status line always says why. When a take comes back silent, ACENTAUR says so in grey with what to try next — no more quiet second and third attempts, and no alarming engine error.

Fixed

• The playhead no longer flickers while ACENTAUR is bypassed; it disappears and comes back at the right place.
• No button changes its size with its state.
• Edges of the surface line up in every host, including LUNA.
• Option-click on a Track Stack track takes the prompt and lyrics you were looking at along to the next track.
• *Again* in Audio Analysis only appears for the source that was analyzed; a new source asks for *Analyze*.

Good to know

• Projects saved with 1.2 need 1.2: ACENTAUR 1.1 opens them without the plug-in's settings. Update every Mac that shares your projects, including a render Mac for ACENTAUR Net.
• The HP Filter is stored with each take and in your project. The same seed with another filter setting gives the same take with a different low end.

What's new

ACENTAUR is now a VST3 plug-in as well. The installer puts an Audio Unit and a VST3 on your Mac, both in the system folders every host scans. They are the same instrument: one surface, the same Modes, the same results. Logic Pro loads the Audio Unit; other hosts may offer either.

One new surface. The plug-in window has been rebuilt from the ground up. The top row holds Capture, the playback menu and the disk symbol for your save location on the left; Presets, Preload, Eject, Help and Settings on the right. The window can be resized by its lower right corner. The status line inside the waveform is the one place ACENTAUR speaks: grey asks, orange works, red objects, green is done.

Text to Music has two types. *Session Player*, the new default, plays one part on the grid of your session — choose a length in bars and a Layer such as Drums, Bass or Guitar, and ACENTAUR plays that part alone. *Composer* is what Text to Music was before: a whole piece from your description, with the Arranger (Off, Small, Medium, Large) and Impact.

Enhance Prompt. A button next to Keep Seed asks the Arranger to rewrite your description in the language the model was trained on, holding tempo, key, meter and length. What you wrote and what is suggested stand side by side; only Apply changes the text, Cancel throws the suggestion away. *Enhance Timing* switches the tempo instruction ACENTAUR adds for Composer on and off.

Track Stack 2.0. The Stack is a multitrack player inside the plug-in. Eight places — Vocals, Backing Vocals, Drums, Bass, Guitar, Keyboard and Synth are fixed, the eighth is yours to choose. Every track carries its own mute and solo, a small fader for listening, its own prompt, lyrics and TAX selection. Accept freezes a take and arms the next track; dragging a track's button drops its WAV into your arrangement.

Cover / Remix is one lane. Instead of two controls and a strength value there is a single lane from COVER to REMIX. At the Cover end the result stays close to the source and re-performs it; toward Remix the written direction takes over. Every position in between makes music.

Presets. In Text to Music a Presets menu offers 50 starter ideas in groups — Drums & Percussion, Bass, Keys & Guitars, Vocals, Brass, Winds & Strings, Synths & Pads, Atmospheres & Textures, Cinematic & FX, Songs & Grooves — followed by your own saved settings and any folders you keep them in, up to three levels deep. Every starter was chosen by ear and brings back exactly that take when your project runs at the tempo in its name. A preset holds description, lyrics, Mode, Model, seed and settings — never audio. Saving a preset keeps the seed of the take you are listening to. If a preset uses a Model that is not installed, the status line names the package to add in Settings.

TAX™ is a page of its own. Open TAX steps in front of the Mode, Back returns you with your text. With a Stable Audio Model only the sound page appears, because those Models do not sing.

Instrumental is a bypass. The switch sits next to the lyrics field. Switched on, the field greys out but keeps showing what is in it, so it is visible what the switch does. Where a Mode already knows the answer, it shows you that answer.

Add Layer always hears the whole capture. The waveform selection is gone from Add Layer; cut in your DAW instead.

The reference is on or off. The Reference Guidance slider has been removed. Cover keeps its own lane.

Audio Analysis runs from Generate. Buttons that make the model work now share Generate's circle-and-arrow: Enhance Prompt, Analyze Audio, and Generate itself, which reads ANALYZE in Audio Analysis.

Drums play without a key. ACE heard the key on every track and pulled drum kits towards it. A drums track now goes out without one, which often sounds clearer. Key and Major/Minor rest while Drums is the target, and the language menu rests wherever the lyrics field does.

The key defaults to C major where music starts from nothing, and a key guess only appears without source audio. A guess is marked as one until you choose.

Faster decoding, no memory spike. The final decoding step of a generation runs in tiles: about 3.7 times faster, and the memory peak at the end of a generation is gone.

Global History shows the seed of every generation, with a copy button.

Keep Capture. A button next to the disk symbol in Cover / Remix, Add Layer and Edit, on by default: the Capture stays the input for every further Generate while the playback menu follows each new Generation. Switch it off to work on the Generation you see; switch it off and on again to make a different Generation the input. Stable Audio Models know no lyrics: with them the lyrics field and the language menu rest.

Settings: deactivate this Mac. Once activated, the license card offers *Manage* and *Deactivate this Mac …*. Deactivating frees the seat in your EINAUX account right from Settings. The Danger Zone lists every destructive action on a line of its own, and the storage location carries a note on drive speed: keep the Models on the internal SSD or a fast external SSD.

Fixed

• LUNA: reloading a session with ACENTAUR on it could crash the host. The surface now outlives the plug-in instance it belonged to.
• LUNA: Capture on a bus or the master with a chosen side-chain recorded the wrong signal. The side-chain is the source whenever the host reports it as chosen — on tracks, buses and the master.
• A Retake played at the current playhead instead of where its original played. It now takes the position of the generation it came from.
• Impact had gone missing from the surface; it is back, next to the Arranger.
• The installer removes the runtime folder left behind by 1.0 installations.
• The ACENTAUR Net pairing question on the render Mac only spoke German. It now follows the language of that Mac.

Under the hood

• One surface for Audio Unit and VST3; the 1.0 window and its layout tests are gone.
• The bill of materials of the packaged runtime and the open-source materials are written from the finished installer at every release and published on the Open Source page.
• Named requirement is macOS 15 or later; the installer still accepts macOS 14.
• The installer asks for agreement to the EINAUX End-User License Agreement and shows the licences of the third-party components ACENTAUR includes.

Update — September 2026

Use as Capture. Keep Capture is no longer a switch: the Capture is always what the next take works from, while the playback menu follows each new Generation. The new push button Use as Capture (Cover / Remix, Add Layer, Edit) makes the Generation you are looking at a Capture of its own, at the same place on your timeline. It appears in the playback menu as a Capture, and the original stays there.

The Capture tells you its key and tempo. After a recording or a drop, the head of the waveform reads, for example, CAPTURE | 16.27s | 118 BPM | F MINOR. A clear key is also placed in the Key menu as a greyed guess — one move corrects it, and a key you chose is never touched. A key ACENTAUR is not sure about carries a question mark and goes nowhere else. Drums alone get none.

Cover / Remix in five steps. The lane now has five fixed positions instead of a continuous run: on the left the result holds on to your source, on the right it is free. Source as Ref, a new button in the Mode row, hands the source to the model as the reference as well.

Analyze Audio does more, and only what you ask for. It now fills in key and language along with description and lyrics. In Cover / Remix and Edit, Analyze Lyrics lets you leave the lyrics run out; in Audio Analysis two switches, Analyze Music and Analyze Lyrics, both on by default, do the same for each half. What is switched off keeps the text that stands in its field.

Longer pieces. Duration in Text to Music now reaches as far as the model renders in one piece — up to 600 seconds at your host tempo (300 bars at 120 BPM) instead of a fixed 64 bars. Stable Audio Models keep 64.

A Generation can be your reference. Drag a Generation from the waveform into the reference slot.

Closer to the reference implementation. New takes start from the same noise an NVIDIA system draws, the arithmetic on the newest Apple chips rounds instead of truncating, and a take with a reference or a source is now repeatable with the same seed. Presets with a fixed seed sound exactly as before.

Fixed

• Ableton Live, Audio Unit: Capture never started unless a sidechain was assigned. It does now. (The VST3 was not affected.)
• Ableton Live, VST3: a file loaded as a Capture before the transport had ever run in that plug-in instance was shown at the right bar but played from bar 1, and every take made from it followed. The import now takes its place from the beat whenever the host has not handed over a position yet. A Capture already loaded that way needs loading again. (Build 20260923.2.)
• A file dropped as a Capture landed late when its sample rate differed from the session's.
• Stopping a Capture before anything was recorded made the previous Capture disappear from the waveform.
• ⌘Z and ⇧⌘Z work in the text fields; ⌘Q and other host commands are no longer held back while ACENTAUR has the focus.
• Generate no longer jumps back to the last Capture; the playback menu lists every Generation of an instance, not only the latest 64.
• Language and key travel with every take, shown as greyed guesses where ACENTAUR guessed them.

Good to know

• Cover / Remix is at its best on sections rather than whole songs.
• On the free (right) positions of the lane the key in the menu decides the harmony — glance at the guess before you generate.
• ACENTAUR Net: update the render Mac and the studio Mac together.
• Projects saved with this update need this update: an earlier ACENTAUR 1.1.0 opens them without the plug-in's settings. Update every Mac that shares your projects.

What's new

Acentaur Settings is an app window now. Acentaur Settings no longer opens a tab in Safari. It has its own window that appears immediately, shows the ACENTAUR mark while the engine starts and then the Settings page. A second click on the app brings that window to the front instead of opening another tab, and closing the window leaves the engine running, so a download continues. If the engine stops, the window says so and starts it again. Settings, Global History and Audio Analysis are tabs at the top of the window; only "Manual" opens the browser. The Settings button in the plug-in opens that window too, and the Settings pages are no longer reachable from a browser at all. Like any app window, it can be moved by its top edge, and nothing on its pages behaves like selectable web text.

All models are smaller and faster. ACENTAUR now ships the six ACE models as 8-bit versions. In blind listening across all six models there was no audible difference, but the models take about half the disk space (ACE 2B: 17 GB instead of 28 GB, ACE XL: 39 GB instead of 64 GB) and render faster. Existing installations are converted automatically on first start, without a new download.

Music Planning is reliable. A bug in the Planner meant that "Keep last seed" could not reproduce a plan, and switching the Planner silently changed later results. Fixed: the Planner is now as seed-true as the model. Results with the Planner therefore sound different from 1.0.0 — not better or worse, but stable for the first time.

Large is back. The large Planner (4B) is part of the XL package and can be selected again.

Impact. The "Prompt Strength" control is now called "Impact" and runs from 2.5 to 3.5 (default 3.0). Lower values sounded noticeably weaker in blind tests; 3.5 sharpens Small and Medium, 2.5 tames Large.

Carry the prompt over. Hold the Option key while switching modes and prompt and lyrics are copied into the new mode. It does not matter whether Option is held when the menu is clicked or only when the entry is chosen. The same works inside Edit when you switch between Repaint, Retouch and Flow.

TAX™ starts with the instrument. In the tag editor the properties of the chosen instrument now come first; Musical Style and Recording & Production move to the end, on screen as in the finished prompt. Every prompt begins with what it is about, style and production frame it. Musical Style and Recording & Production now live on their own page "Style & Production", switchable at the top next to Back; the sub-pages of Musical Style and the jump to the top when turning the page are gone. The tabs are named like the fields on the main page, "Sound & Interpretation" and "[Performance] & Lyrics", and the instrument menu stays visible on both. On the right, a card holds title, hint, prompt, Clear and Apply. A new row "Arrangement" offers entry marks such as [Band enters], [Drums enter] or [Beat Drop]. The row "Vocal Delivery" now appears only for Vocals and Backing Vocals, at the very top. And the Lyrics page resolves contradictions on the spot: when two marks stand directly after each other with no text in between, a new mark of the same kind replaces the old one — Chorus replaces Verse, High energy replaces Low energy. With text in between they are two moments, and everything stays allowed.

Fixed

• Installing an update right after quitting Logic could fail with macOS' generic “an error occurred while running scripts”: the Acentaur engine was still shutting down in the background and the installer mistook it for another program. The installer now waits for the engine to finish shutting down.
• During a first installation, the Audio Analysis and Detailed Description packages could not be selected: their cards showed no size and the Install button stayed greyed out. They were tied to the license, and nobody is activated yet at that point. They can now be installed right away; using the analysis still requires an activated license.
• Activating with a fresh e-mail code could take longer than the ten minutes the activation request was valid for. The request now stays valid for an hour, and if a confirmation does arrive too late, the account page says so instead of looking like a success.
• A Settings page left open from an earlier engine session kept showing progress but silently ignored every click. Such a page now says that it belongs to an earlier session and locks its buttons. With the new Settings window this can no longer happen in the first place.

• Clicks in the tag window (TAX) landed on the main page behind it.
• The seed field showed "Auto" as text; a pasted seed ended up inside the word. "Auto" is now a placeholder and the field is empty.
• A red error message (for example about the seed) was overwritten by the next control movement.
• Edits (Repaint, Retouch, Flow) loaded a Planner they never used; on Macs without an installed Planner they failed.
• Cover without a given key lost the automatically detected key.
• If the model service dies in the middle of a job, ACENTAUR restarts it and repeats the job instead of showing an error.
• The Settings icon had gone missing; the Engine now has an icon of its own.
• Global History: coming back from Finder or Logic jumped the list to the top; position, selection and playback are now preserved.
• Global History showed internal names such as "lego"; it now shows the names from the plug-in (Add Layer, Retouch, Track Stack, …).
• A click on the empty area of the tag window (TAX) made the plug-in window deaf to all input until Logic was clicked elsewhere, which gave back exactly one click. Focus now stays in the window.
• Capture with an assigned sidechain ended right after a single audio block: the noise floor of the main input counted as signal and won against the still-silent sidechain bus. An assigned sidechain is now always the source, and only what rises above minus 80 dB counts as signal.
• Cmd+Z in a prompt field went to Logic and undid the last step there — in the worst case the insertion of ACENTAUR itself. The text fields now have their own undo and redo (Cmd+Z, Shift+Cmd+Z), including the lyrics field in TAX.
• After reopening a project, the status line showed the result line of the last session although Generate was locked. It now says what is needed next, for example "Briefly press Play in your DAW".
• Capture could start on its own as soon as audio reached the input while the transport stood still, for example from a microphone with input monitoring or from a rendering side chain. Capture now waits for Play in your DAW and audio together, and it ends when your DAW stops.

Under the hood

• Setup now states the memory the 8-bit models actually need: ACE XL from 16 GB (XL Turbo), 32 GB for all models and the Large planner; ACE 2B from 16 GB, 24 GB recommended. Nothing is blocked; the figures are guidance.

• The Planner is used only in Text to Music and is unloaded when you leave the mode (memory).
• Conditioning runs in fp16 (no audible difference, 31 of 31 identical).
• Model service logs are rotated at 8 MB.
• ACENTAUR Net jobs are deleted after they are collected, leftovers after 24 hours.
• No more unused folders in the application data folder.
• Setup texts: storage requirements updated, Large listed in the XL description.
• A damaged installation now says "Installation damaged · reinstall Acentaur" in the status line instead of misbehaving.
• Third-party notices now name libsndfile (LGPL-2.1-or-later), which ships inside the soundfile module, next to soxr.

Known and open

• Eight bars sometimes sound like six or seven. The file always has exactly the ordered length at the DAW's tempo and starts on the one. But the model composes an ending and in most runs leaves the last one or two bars empty. Until ACENTAUR trims this automatically: order two bars more and shorten the region in the DAW, or work with 4-bar phrases, which end cleanly far more often. Affects Text to Music and the first layer of an empty Track Stack.
• On 16 GB Macs, XL Turbo is feasible according to simulation but tight; acceptance on a real 16 GB Mac is pending.
• In tests, the model service occasionally quit in the middle of a job; the cause is still open, the update catches it.