gary4juce

gary4juce

gary v5: yue2 and a native stem separator are now inside the DAW.

a VST3/AU plugin for musicians who want AI to meet them where they actually live. for me, that’s ableton. for you, it might be fl studio or, if you’re insane, reaper lol

https://thepatch.gumroad.com/l/gary4juce

latest stable releases:

• gary4juce v5.0.2 (windows VST3 + standalone)
• gary4juce v5.0.0-mac (macOS AU/VST3 + standalone)

recommended local companions:

• windows: gary4local v0.6.0
• macOS: gary4local mac v0.4.0

localhost backend source repos:

• windows: gary4local
• macOS: gary4local mac

videos about it here when i can: https://youtube.com/@thepatch_dev

---

built for musicians

If you just want pure text-to-music, you probably do not need a VST. This project exists for people who want AI to sit inside the session with them.

gary4juce now gives you eight AI music models directly in your DAW:

• yuey (YuE2) - songs with or without lyrics in your tempo and key, remixes of your audio through its own transcription, continuations, and a score you can edit and drag in as MIDI
• sa3 (stable-audio-3) - text-to-audio, loops, transforms, continuations, LoRA blending, seed recall, key/BPM-aware prompting
• gary (musicgen) - continuation/anti-looper. Extends your audio in creative directions
• jerry (stable-audio-open-small) - BPM-aware 12-second loop generation in under a second
• rc-jerry (foundation-1) - BPM and key-aware 4/8-bar loop generation with structured prompt assembly
• carey (ace-step) - stem generation, audio continuation, and remix/cover with lyrics and multilingual support
• terry (melodyflow) - audio transformation. Turn your guitar into an orchestra
• darius (magenta-realtime) - high-quality 48 kHz continuations with style control (probably deprecating this one soon)

and one thing that isn’t a model you prompt: a stem separator, built on stems.cpp, that splits gary’s output or your recording buffer into stems on your own machine.

Put it on your master, press play, record some audio, and start iterating.

yuey runs on the remote backend for everyone, and locally through gary4local v0.4.0 on Windows and macOS.

---

latest update

v5.0.2 - 10/31

Carey moves to ace-server. gary4local v0.6.0 runs Carey on acestep.cpp, and Carey’s tabs speak its job API on localhost and the remote backend alike. there’s a new generate tab that makes audio from a caption; continue, cover, lego and complete work on your recording as before, with your LoRAs, progress and cancel. on localhost, each tab’s advanced settings choose 2B or XL; the remote backend serves XL. lego and complete are still hit or miss, and their tabs say so.

migrate Carey first. local Carey needs the C++ runtime, so press migrate runtime on Carey in gary4local v0.6.0 before using it on localhost. the remote backend serves both the old and new API, so an older plugin keeps working there.

the saved recording keeps its own waveform lane while an unsaved take is present. playback and stem separation use the saved buffer, opening a file restores it there, and new recording audio shows as unsaved from the first sample.

recommended local companions: gary4local v0.6.0 on Windows, released alongside this one, and gary4local mac v0.4.0 on macOS. the stem separator runs inside gary4juce and doesn’t need either.

older release notes now live in docs/CHANGELOG.md.

---

roadmap

• bring local yuey to macOS (gary4local mac v0.4.0)
• unify gary4local into one cross-platform app through the native gary4local migration
• introduce optional lightweight native stem separator
• revisit Carey complete mode so it can do the upstream-style accompaniment workflow (if it actually works like it should…)

---

installation

windows

The Windows release has two downloads: the VST3 ZIP for use in a DAW and the standalone ZIP for running gary4juce as its own app.

For the VST3:

1. Close your DAW.
2. Use Extract All on the ZIP.
3. Choose C:\Program Files\Common Files\VST3\ as the destination, or extract somewhere else and copy the gary4juce.vst3 folder there.
4. Reopen your DAW and rescan plugins if needed.

You can put the VST3 literally anywhere as long as your DAW scans that location. C:\Program Files\Common Files\VST3\ is just the default path most DAWs already check.

For the standalone app, extract the standalone ZIP anywhere you want to keep it, then run gary4juce.exe from the extracted gary4juce-standalone folder. If Windows reports missing Visual C++ runtime DLLs, run the included, Microsoft-signed VC_redist.x64.exe once and try again. License and source information for gary4juce lives in the adjacent Legal folder.

If permission errors appear, run Command Prompt as admin:

xcopy "path\to\extracted\gary4juce.vst3" "C:\Program Files\Common Files\VST3\gary4juce.vst3" /E /I /Y

LMMS support is not working yet. We did an initial VST2/LV2 compatibility pass and documented the exact Windows alpha environment here: LMMS compatibility notes.

macOS AU/VST3/standalone

The macOS release has separate universal DMGs for AU, VST3, and the standalone app.

1. Quit your DAW and the standalone app.
2. Open the DMG for the format you want.
3. Drag to the matching folder:
   • gary4juce.component -> Components (Audio Unit)
   • gary4juce.vst3 -> VST3
   • gary4juce.app -> Applications
4. Reopen your DAW and rescan if you installed a plugin.

GarageBand and Logic use AU. Ableton, FL, Reaper, Cubase, and Bitwig can use VST3.

---

backends

The plugin can use either:

• remote backend - self-hosted on a DGX Spark, free for now, limited to the models I have loaded
• localhost - your machine, requires GPU, full control

Remote base URL: https://g4l.thecollabagepatch.com

gary4local

Use the dedicated apps for localhost:

• windows: gary4local
• macOS: gary4local mac

They manage local envs for gary, terry, jerry, carey, foundation-1, and SA3, and on Windows a native runtime for yuey. Model coverage varies by platform, but SA3 is available in both companion apps.

the stem separator isn’t one of these services. it runs inside the plugin.

Recommended hardware:

• 10 GB+ GPU VRAM minimum
• 16 GB+ recommended for heavier local models
• 24 GB+ recommended for Darius-style separate backends

SA3 backend

SA3 can run on the remote backend:

https://g4l.thecollabagepatch.com/sa3

or the local gary4local service on Windows/macOS:

http://localhost:8006

Public upstream repo: https://github.com/stability-ai/stable-audio-3

Darius backend

Darius is too heavy to run alongside the other gary4local services.

Easiest path: duplicate this Hugging Face Space:

https://huggingface.co/spaces/thecollabagepatch/magenta-retry

Use an L40s or A100-class runtime. Enter the duplicated Space URL in the Darius tab.

Local Docker option:

git clone https://github.com/betweentwomidnights/magenta-rt
cd magenta-rt
docker build -f Dockerfile.cuda -t magenta-rt .
docker run --gpus all -p 7860:7860 magenta-rt

---

how to use

1. Put gary4juce on your master track, a bus, or any track you want to listen to.
2. Press play in your DAW.
3. Save the recording buffer when you have audio you want a model to react to.
4. Generate, transform, continue, crop, drag, and repeat.

yuey tab

yuey uses YuE2.

• create - a song in your tempo and key, with lyrics or instrumental. by default it renders over a chord scaffold straight away; “let yuey plan” lets the model compose the whole score first
• remix - transcribe your recording into a score and render it again in a new style
• continue - extend your audio from its score, or from the audio itself

the prompt is shared across all three, so switching sub-tabs keeps what you typed.

choose midi in remix or continue to drop or pick a melody file, a chords file, or both. .mid, .midi and saved MIDI Live Clips (.alc) are supported. crop to whole bars and give both lanes the same length; Gary imports the notes, not the clip’s instrument or effects. the guide explains the supported rhythms and chord shapes.

every render keeps its score: drag it into your DAW as MIDI, or open the score window to edit it, transpose it, or halve or double the tempo, and render it again. the score window has the same prompt box and dice, so you can change the sound without leaving it.

full guide: YUEY.md

learn more: https://github.com/multimodal-art-projection/YuE

jerry tab

The Jerry tab now has three sub-tabs:

sa3 - stable-audio-3

• Generate text-to-audio up to 380 seconds on localhost or 240 seconds remotely.
• Toggle loop mode for 4/8/16-bar loop generation.
• Transform either the saved recording buffer or the current output audio.
• Continue either the saved recording buffer or the current output audio.
• Use optional key/scale and automatic DAW BPM prompting.
• Use seed recall for reproducible takes.
• Use LoRA sliders in advanced settings. Sliders default to 0.
• Use dice prompts from the default pool, or from selected LoRA pools when one or more LoRA sliders are above 0.

Full guide: SA3.md

jerry - stable-audio-open-small

• Generates short loops aligned to your DAW BPM.
• Smart loop mode can bias toward drums or instruments.
• Localhost supports custom finetunes through the model picker.

Learn more: https://huggingface.co/stabilityai/stable-audio-open-small

rc-jerry - foundation-1

• Generates 4 or 8 bar loops synced to BPM and key.
• Uses structured prompt assembly through knobs, toggles, and tag controls.
• Randomize builds coherent presets through the backend prompt engine.
• Presets can be saved/loaded as .f1preset files.
• Available on remote, gary4local windows, and gary4local mac on Apple silicon.

Learn more: https://huggingface.co/RoyalCities/Foundation-1

gary tab

Gary uses MusicGen continuation models.

• Uses the first selected seconds of your recording buffer.
• Generates continuation audio.
• Output can be cropped, continued, retried, or dragged to the DAW timeline.

Model lists are fetched dynamically from the backend, so the menu depends on what is loaded locally or remotely.

Learn more: https://github.com/facebookresearch/audiocraft

carey tab

carey uses ACE-Step.

• lego - generate vocals/backing vocals over your audio
• complete - extend audio into a full continuation
• cover - remix/restyle with caption guidance

there used to be an extract mode here. it’s hidden as of v5.0.0 because the built-in stem separator does that job better.

shared lyrics editor, 50-language support, key/scale/time signature selection, caption popouts, and LoRA support live here.

for lego mode, regular acestep-v15-base is still the safer default if you don’t have a LoRA trained. xl-base gets much more interesting when a matching xl-base LoRA is attached, especially for vocals and backing vocals. some instrument bleed can happen, but it often fits the source audio anyway.

full guide: CAREY.md

learn more: https://github.com/ace-step/ACE-Step-1.5

terry tab

• Transforms audio with style presets or custom prompts.
• Can use either the recording buffer or current output audio.
• Undo is available after transforms.

Learn more: https://huggingface.co/spaces/facebook/Melodyflow

darius tab

• High-quality 48 kHz continuations.
• Style steering with prompts, base model, or custom weights.
• Works on the recording buffer or current output audio.
• Requires a separate backend.

Learn more: https://github.com/magenta/magenta-realtime

stem separator

this one isn’t a tab. it’s a stems handle at the bottom right of the output waveform, and of the recording buffer once it has audio. it stays dimmed until you’ve installed the runtime and at least one model. click it then and it opens the stem separator settings, which also live in the settings menu under stem separator.... the runtime is small (about 19 MB on windows, 6 MB on mac). the models download from Hugging Face.

• pick a model, press separate. every stem gets its own waveform you can play, pause and seek.
• drag the stems you want into your DAW. only a stem you drag out is kept. the rest are deleted when you close the popup.
• drop a stem on the recording buffer to replace what’s in it. that’s how you pull the drums out of something you dragged in from the DAW.
• with mdx23c_drumsep downloaded, a drums stem gets a split kit button that splits it into kick, snare, toms, hi-hat, ride and crash, right there in the popup. pick DrumSep from the menu to split the audio itself instead, which is what you want for a drum loop. a from menu beside it switches between the two.
• it’s disabled while a generation is running.

you can now drop or pick up to ten minutes of audio into the recording buffer, regardless of the selected model or local/remote connection, then separate the whole track. larger files open the selection window so you can choose a section up to ten minutes. double-click the buffer to select a shorter section for generation: the window opens at the active model’s input limit, and the handles cannot stretch past it (30 seconds on Terry; 380 on local SA3/Carey or 240 remotely). an orange hint shows when the buffer exceeds the selected model’s input limit; oversized requests are stopped before upload. importing a full track does not raise a generator’s duration limit.

the buffer grows for long imports when needed; live recording keeps its existing five-minute limit. separation still computes all the model’s stems. none of the models gets faster from choosing fewer outputs.

it runs on the GPU by default (Vulkan on windows, so NVIDIA, AMD and Intel all work, and Metal on apple silicon), with a toggle for the CPU. on the CPU, htdemucs and the UVR cleanup models are the ones to use. the RoFormers are much slower there, and DrumSep runs at about a tenth of realtime. intel macs are CPU only.

model	stems	download
stems
htdemucs	drums, bass, other, vocals	100 MB
htdemucs_6s	drums, bass, other, vocals, guitar, piano	71 MB
htdemucs_ft	same four as htdemucs, best quality, about 4x slower	401 MB
vocals and instrumental
mel_band_roformer_kim	vocals, instrumental	436 MB
bs_roformer_viperx_317	vocals, instrumental	610 MB
mdx_net_kim_vocal_2	vocals, instrumental; small and quick	64 MB
noise, echo and reverb
uvr_denoise	noise, denoised	121 MB
uvr_denoise_lite	noise, denoised; smaller and quicker	17 MB
uvr_deecho_normal	no_echo, echo	121 MB
uvr_deecho_aggressive	echo, no_echo; removes more	121 MB
uvr_deecho_dereverb	no_reverb, reverb; echo and reverb together	213 MB
bs_roformer_dereverb_room	noreverb, reverb; for vocals recorded in a room	112 MB
drum kit
mdx23c_drumsep	kick, snare, toms, hh, ride, crash; feed it drums	417 MB

the cleanup models each give you both halves, the clean audio and what was taken out of it, so you can keep either. the two de-echo models list them in opposite orders because that’s how UVR trained them.

htdemucs is the default. bs_roformer_viperx_317, mdx_net_kim_vocal_2, mdx23c_drumsep and the five uvr_ models have no license upstream, and bs_roformer_dereverb_room is GPL-3.0, so check the license section before you use them for anything beyond your own sessions.

learn more: https://github.com/betweentwomidnights/stems.cpp

---

building from source

Requirements:

• JUCE 8.0.8
• Visual Studio 2022 on Windows
• Xcode on macOS

Steps:

1. Open gary4juce.jucer in Projucer.
2. Save the project to regenerate build files.
3. Open the generated IDE project.
4. Build release configuration.

Maintainers: see the release checklist for packaging and verification.

---

community

discord: https://discord.gg/VECkyXEnAd

musicgen community: https://discord.gg/Mxd3nYQre9

email: [email protected]

twitter/x: https://twitter.com/@thepatch_kev

---

credits

• stable-audio-3: Stability AI (repo)
• underfit: dadabots (repo)
• musicgen: Meta AI / Audiocraft team
• stable-audio-open-small: Stability AI
• foundation-1: RoyalCities (model, tools)
• ace-step: ACE-Step team (repo)
• melodyflow: Meta AI / Audiocraft team
• magenta-realtime: Google Magenta team
• stems.cpp: tinycrops (repo), the native runtime behind the stem separator
• htdemucs: Meta AI (demucs)
• mel-band-roformer: KimberleyJSN (model)
• bs-roformer ep_317: viperx, released through the UVR model repo
• JUCE: JUCE framework
• community finetunes: lyra, vanya, hoenn, CZ-84, and everyone contributing models

Special thanks to Zach and the guys at Stability for letting me be part of the Stable Audio 3 beta while this integration came together.

---

license

The source code and other original material in this repository, including earlier gary4juce versions authored by the copyright holder, are free software licensed under the GNU Affero General Public License v3.0 only (AGPL-3.0-only). You may use, study, modify, and redistribute that material under the license. Redistribution and network use of modified versions are subject to the AGPLv3’s notice and Corresponding Source requirements.

Copyright © 2025-2026 Kevin Griffing. Developed and published by the collabage patch, inc.

This project is built with JUCE 8.0.8 under JUCE’s AGPLv3 option. See Third-Party Notices for the pinned JUCE source and license information.

This license statement is limited to this repository. It does not set the license for the separate gary4local applications or installers, model weights, hosted services, or any other separately distributed part of the broader Gary ecosystem. Consult each project’s own license before using or redistributing it.

The AI models and backend services used with gary4juce are separate works and are not included in this repository. Code and model weights may use different licenses:

• stable-audio-3: MIT code; weights use the Stability AI Community License
• musicgen: MIT code and CC-BY-NC-4.0 weights
• stable-audio-open-small: Stability AI Community License
• foundation-1: Stability AI Community License
• ace-step 1.5: MIT
• melodyflow: MIT code and CC-BY-NC-4.0 weights
• magenta-realtime: Apache-2.0 code; model-weight terms depend on the selected version
• stems.cpp: MIT
• htdemucs: MIT (Meta)
• mel-band-roformer (Kim): MIT
• bs-roformer viperx ep_317: no license upstream. it’s published as “other” with a notice saying so. treat it as unlicensed.

Always consult the exact upstream code and model version before commercial use or redistribution.

---

support the project

If gary4juce is useful to you:

• share your creations and tag @thepatch_kev
• contribute finetunes to the community
• help improve docs
• Gumroad: https://thepatch.gumroad.com/l/gary4juce (this is one of the only ways to support the project monetarily rn…working on some easier paths like github sponsors)