gary v4 is stable: stable-audio-3 is 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:
recommended local companions:
- windows: gary4local v0.1.18
- macOS: gary4local mac v0.1.11
localhost backend source repos:
- windows: gary4local
- macOS: gary4local mac
videos about it here when i can: https://youtube.com/@thepatch_dev
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 seven AI music models directly in your DAW:
- 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, extraction, 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
Put it on your master, press play, record some audio, and start iterating.
SA3 now runs on the remote backend, the Windows gary4local SA3 service, and gary4local mac.
v4.0.4 opens up Carey LoRA usage in lego mode and refreshes the Carey docs around what has worked best in testing.
lego mode is still usually more reliable and musical with regular
acestep-v15-base when no LoRA adapter is loaded. xl-base gets much more
interesting once a matching xl-base LoRA is attached, especially for vocals and
backing vocals.
this release is paired with gary4local v0.1.18, which fixes local Carey seed reporting and updates ACE-Step LoRA captioning so LM-hallucinated lyrics are left out of training sidecars.
older release notes now live in CHANGELOG.md.
- add SA3 to gary4local on Windows
- add SA3 to gary4local mac on macOS
- ship local SA3 training on Windows and macOS using underfit as the source of truth
- clean up and release a proper standalone app
- ship local ACE-Step LoRA training in gary4local on Windows using Side-Step as the reference point
- revisit Carey complete mode so it can do the upstream-style accompaniment workflow
- enable the Carey
xl-sftmodel on the remote backend
- Close your DAW.
- Use Extract All on the ZIP.
- Choose
C:\Program Files\Common Files\VST3\as the destination, or extract somewhere else and copy thegary4juce.vst3folder there. - 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.
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 /YLMMS support is not working yet. We did an initial VST2/LV2 compatibility pass and documented the exact Windows alpha environment here: LMMS compatibility notes.
- Quit your DAW.
- Open the DMG.
- Drag to the matching folder:
Gary4Juce.component-> Components (Audio Unit)Gary4Juce.vst3-> VST3
- Reopen your DAW and rescan.
GarageBand and Logic use AU. Ableton, FL, Reaper, Cubase, and Bitwig can use VST3.
The plugin can use either:
- remote backend - my server, free, on a spot VM, limited to the models I have loaded
- localhost - your machine, requires GPU, full control
Remote base URL: https://g4l.thecollabagepatch.com
Use the dedicated apps for localhost:
- windows: gary4local
- macOS: gary4local mac
They manage local envs for gary, terry, jerry, carey, foundation-1, and SA3. Model coverage varies by platform, but SA3 is available in both companion apps.
Recommended hardware:
- 10 GB+ GPU VRAM minimum
- 16 GB+ recommended for heavier local models
- 24 GB+ recommended for Darius-style separate backends
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
Local SA3 training source of truth: https://github.com/dada-bots/underfit
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- Put gary4juce on your master track, a bus, or any track you want to listen to.
- Press play in your DAW.
- Save the recording buffer when you have audio you want a model to react to.
- Generate, transform, continue, crop, drag, and repeat.
The Jerry tab now has three sub-tabs:
- Generate text-to-audio up to 300 seconds.
- 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
- 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
- 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
.f1presetfiles. - Available on remote, gary4local windows, and gary4local mac on Apple silicon.
Learn more: https://huggingface.co/RoyalCities/Foundation-1
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 uses ACE-Step.
- lego - generate vocals/backing vocals over your audio
- complete - extend audio into a full continuation
- cover - remix/restyle with caption guidance
- extract - attempt target stem extraction from your recording buffer
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
- 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
- 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
gary4juce gets better as more finetunes exist.
Train through Audiocraft:
https://github.com/facebookresearch/audiocraft
As of late 2025, Google Colab is painful for Audiocraft training due to dependency conflicts. Local training is the practical path.
Train with stable-audio-tools:
https://github.com/Stability-AI/stable-audio-tools
Encode, train, select checkpoints, upload to Hugging Face, then load through the Jerry localhost finetune picker.
Local SA3 training for both gary4local companions now uses dadabots' underfit as the source of truth.
SA3 LoRA workflows are new and still settling. The v4 plugin UI is already shaped around multi-LoRA strength sliders and LoRA-aware dice pools so the backend can grow into that workflow cleanly.
Upstream repo: https://github.com/stability-ai/stable-audio-3
Magenta Realtime is one of the friendlier finetuning paths:
https://github.com/magenta/magenta-realtime
Upload weights to a Hugging Face model repo and point the Darius tab at it.
gary4juce/
+-- Source/
| +-- PluginProcessor.cpp/h
| +-- PluginEditor.cpp/h
| +-- Components/
| | +-- Gary/GaryUI.cpp/h
| | +-- Jerry/JerryUI.cpp/h
| | +-- Jerry/SA3UI.cpp/h
| | +-- Foundation/FoundationUI.cpp/h
| | +-- Carey/CareyUI.cpp/h
| | +-- Terry/TerryUI.cpp/h
| | \-- Darius/DariusUI.cpp/h
| \-- Utils/
| +-- Theme.h
| +-- IconFactory.cpp/h
| \-- BarTrim.cpp/h
+-- CAREY.md
+-- SA3.md
+-- docs/
\-- gary4juce.jucer
Requirements:
- JUCE 8.0.8
- Visual Studio 2022 on Windows
- Xcode on macOS
Steps:
- Open
gary4juce.jucerin Projucer. - Save the project to regenerate build files.
- Open the generated IDE project.
- Build release configuration.
- SA3 launch notes: output loudness and continuation tails are still being tuned.
- SA3 local backend: available through gary4local on Windows and gary4local mac.
- Windows Defender: not codesigned, so Windows may complain.
- Darius hardware: 24 GB+ VRAM is strongly recommended.
- Terry variability: Melodyflow is experimental and can be wonderfully strange.
- Carey complete mode: useful, but the upstream-style accompaniment workflow still needs a proper UI/backend pass.
discord: https://discord.gg/VECkyXEnAd
musicgen community: https://discord.gg/Mxd3nYQre9
email: kev@thecollabagepatch.com
twitter/x: https://twitter.com/@thepatch_kev
- 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
- 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.
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 (C) 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
Always consult the exact upstream code and model version before commercial use or redistribution.
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)