Stability Matrix: One Installer for ComfyUI, Forge and InvokeAI

Stability Matrix is a free desktop app that installs and updates local image-generation tools for you. It installs ComfyUI, InvokeAI, Forge, SwarmUI, SD.Next and two dozen more from one window, gives each its own Python environment, and lets them all share one models folder. It runs on Windows, macOS (Apple Silicon) and Linux, and it is made by Lykos AI.

This guide covers what it installs, how it shares models between packages, where it keeps its data, what it sends back to Lykos, and the licence. It also lists how actively each package it offers is still maintained, because the app still offers several tools that have stopped shipping. BFWR has not run Stability Matrix. Everything here comes from its source code, documentation, release history and legal pages, plus the GitHub API for each package, all read on 1 November 2026.

Quick answer

• What it is: a package manager for local AI image and video tools. The current release is v2.16.5 (21 October 2026). The source code is AGPL-3.0, but the downloadable builds come under a separate Lykos EULA.
• What it installs: 30 packages are defined in the source, and 29 are offered across Inference, Training and Legacy tabs. One, SimpleSDXL, is hidden because of a critical security advisory.
• Maintenance: Stability Matrix does label stale tools as Legacy, including lllyasviel’s original Forge. But AUTOMATIC1111 and reForge still sit in the main Inference tab, and neither has had a commit in the last 90 days. reForge’s own README now recommends other forks “for stability”.
• Models: one shared Models folder, linked into each package by symlinks (junctions on Windows) or by writing the package’s own config file. ComfyUI uses the config method.
• Privacy: no account is needed to install packages or browse models. Usage analytics are offered at first launch with sharing as the default button, and crash reporting is on unless you start the app with --no-sentry.

What Stability Matrix is

The project is LykosAI/StabilityMatrix on GitHub, started in June 2023, with about 9,000 stars. Its README calls it a “Multi-Platform Package Manager and Inference UI for Stable Diffusion”. The Inference UI is a built-in generation screen that runs on an installed ComfyUI.

Builds come from the Lykos downloads page or GitHub Releases. Windows 10 and 11 get a .zip (extract it and run StabilityMatrix.exe), macOS gets a .dmg for Apple Silicon only, and Linux gets a .zip with an AppImage, plus an AUR package for Arch.

# Linux, from the official installation docs
unzip StabilityMatrix-linux-x64.zip
chmod +x StabilityMatrix.AppImage
./StabilityMatrix.AppImage

Some distributions also need libfuse2. The AUR build “cannot update itself through Stability Matrix’s in-app updater”, so it updates through your system package manager.

Every package it can install

Each package is a C# class in StabilityMatrix.Core/Models/Packages/. Thirty register as installable, and the Add Package screen sorts them into three tabs by their declared type:

• Inference (13): AUTOMATIC1111’s WebUI and its DirectML fork, ComfyUI, ComfyUI-Zluda, Forge Neo, reForge, InvokeAI, SD.Next, SwarmUI, Cogstudio, FramePack, FramePack Studio and Wan2GP.
• Training (5): AI-Toolkit, OneTrainer, kohya_ss, FluxGym and Fizgig.
• Legacy (11 shown): the original Forge, Forge Classic, AMDGPU Forge, four Fooocus variants, RuinedFooocus, SD Web UI-UX, SDFX and VoltaML.

The thirtieth, SimpleSDXL, is never shown. Its class carries a critical advisory, “Undisclosed Data Collection and Remote Access in simpleai_base Dependency”, and the package factory drops any package with a known vulnerability. Packages that do not support your GPU are hidden by default too, so a Mac or AMD machine shows fewer.

How well maintained each package is

The docs describe Legacy as “Older packages that are maintained for existing users but not recommended for new installations. May be stale and no longer receiving updates.” To see whether that tab catches every stale package, we read each upstream repository through the GitHub API. “Last commit” is on the branch Stability Matrix installs, the 90-day window starts on 3 August 2026, and 300+ means the count stopped at 300.

None of these repositories is archived. What the numbers show:

• The Legacy tab is mostly right. lllyasviel’s Forge (last commit June 2025), Fooocus, VoltaML and SDFX are there. Our image editor roundup found the same Forge dormancy in September.
• AUTOMATIC1111 is not in it. A1111 sits in the Inference tab with no disclaimer. Stability Matrix’s docs say its dev branch “is installed by default as it is in active development while the” main branch “has been in a stale state.” That branch’s last commit was on 2 March 2026, its fourth in a year. And the install code does not pick dev: A1111 publishes releases, so the installer starts in release mode and installs v1.10.1 from February 2025.
• reForge is ranked near the top. The docs call it “A fast-moving Forge fork”. Its last commit was on 14 April 2026, and its README now opens with “Suggested repos instead of reForge for stability”, pointing to Forge Classic and Forge Neo. The source ranks it Recommended, so on an NVIDIA PC the one-click installer lists it second, after Forge Neo and ahead of ComfyUI.
• Three more Inference packages had no commits in 90 days: FramePack, Cogstudio’s CogVideo repo and the DirectML fork. The DirectML class also inherits a dev branch from the A1111 class that the renamed upstream repo no longer has.
• One Legacy package is active. RuinedFooocus had 26 commits and a release on 31 October. It is in Legacy only because its class inherits that type from Fooocus.

For a new install, the inference packages that are clearly still moving are ComfyUI, InvokeAI, SwarmUI, SD.Next, Forge Neo and Wan2GP. If you want the A1111-style interface, Forge Neo is the maintained route.

The shared models folder

All packages read from one library, the Models folder in the data directory, with subfolders such as StableDiffusion, Lora, VAE and TextEncoders. Each installation uses one of three sharing methods:

• Symlink: the package’s model folders become links. The docs say “Stability Matrix uses directory junctions on Windows and symbolic links on other platforms.” It first tries to move existing files into the shared folder. A1111, its forks and SD.Next default to this.
• Config: the shared paths are written into the package’s own config file. ComfyUI defaults to this, through extra_model_paths.yaml, as do InvokeAI, SwarmUI and Fooocus.
• None: no sharing. Most training packages use this.

With symlinks, “deleting a model through a linked folder deletes the same file used by other packages.” Sharing also does not make a file compatible: an SD 1.5 package can see a FLUX checkpoint and still fail to load it. Our guide to checkpoints, LoRAs and VAEs explains what goes where. Outputs are shared into an Images folder by default; turn that off and a package’s images are deleted when you uninstall it.

Data directory and portable mode

Everything lives in one data directory: Packages (each with its own venv), Models, Images, Assets, Workflows and settings.json. Portable mode is on by default at setup and puts a Data folder next to the app, so the whole setup moves as one folder. Otherwise the defaults are %AppData%\StabilityMatrix on Windows, ~/StabilityMatrix on Linux and ~/.config/StabilityMatrix on macOS. Changing the location later “does not move your existing packages, models, or images for you.” FAT32 and exFAT drives are not supported.

Python and Git

The README lists “Embedded Git and Python dependencies, with no need for either to be globally installed”. The app downloads uv, Git and each package’s Python version into Data/Assets, creates a separate virtual environment per package, and installs the PyTorch build that matches your GPU: CUDA, ROCm, DirectML, ZLUDA, Intel IPEX, Apple MPS or CPU. Two tools that need different PyTorch versions cannot break each other. The cost is disk: one package is “commonly in the multi-gigabyte range” before any models. For what your GPU can run, see our image generation VRAM calculator.

Model browser and accounts

The Model Browser searches CivitAI and Hugging Face and saves each download, with metadata and a preview, into the right folder. Browsing and downloading ordinary models needs no account. Two optional keys add more: a CivitAI key “Enables downloading early-access models from CivitAI and browsing NSFW-tagged models”, and a Hugging Face token “Enables downloading gated models (e.g., FLUX) from the Hugging Face Hub.” The docs say keys are stored encrypted, tied to the machine. A Lykos account is only needed for memberships, the Preview and Dev channels and cloud features such as Prompt Amplifier.

What it sends back to Lykos

Usage analytics. First launch asks whether to share analytics. In the source, the buttons are “Don’t Share Analytics” and “Share Analytics”, and the second is the default. The settings docs say “Analytics are enabled by default (opt-out)”: package installs, first-run choices and a daily launch ping with the app version and OS. Turn them off in Settings, System, Analytics.

The documents disagree on hardware data. The settings docs say “No hardware information, model browsing history, personal data, or API keys are included.” The Lykos EULA’s data list includes “Operating system and GPU information”, sent to auth.lykos.ai, and says nothing is sent when analytics are off.

Crash reporting. The app starts Sentry error reporting, with session tracking, before its window opens. Neither the settings docs nor the settings code offer a switch. The privacy policy says “This can be disabled via launch argument”, and the argument is --no-sentry.

The packages themselves are separate programs. The EULA says Lykos is not responsible for “any data collection, processing, or transmission performed by third-party software.”

Updates and release cadence

The app checks a signed manifest at cdn.lykos.ai, and “updates are never installed without confirmation.” The stable channel is free; “Preview and Dev channels require an active Lykos membership.” Releases are steady: v2.16.0 on 9 June 2026, then five point releases to v2.16.5 on 21 October, and at least 100 commits to main in the last 90 days. Packages update separately, each with its own Check for Updates and Change Version options, so you can roll one back if an upstream change breaks a workflow.

The licence: AGPL source, EULA binaries

The GitHub source is AGPL-3.0. For personal use that barely matters: you can run, study and modify it. The extra obligations apply only if you distribute a modified version, or let others use one over a network; then you must offer them your source.

The README adds a second layer: “Binaries and executable releases are licensed under the End User License Agreement”. The EULA (26 March 2026) covers builds from lykos.ai, GitHub Releases and the auto-updater. It says “This EULA doesn’t limit your open-source rights”, but for the binaries it bars redistributing modified builds under the Stability Matrix or Lykos names, and says you may not “Charge money, fees, or demand compensation for distributing the Application”. Lykos claims no ownership of what you generate. Each package keeps its own licence, from MIT to AGPL-3.0, and models add a third layer, which our model licence checker covers. This is our reading of the published texts, not legal advice.

FAQ

Is Stability Matrix free?

Yes, including the stable update channel. Paid memberships add Preview and Dev builds and credits for cloud features.

Does Stability Matrix still install AUTOMATIC1111 and Forge?

Yes. A1111 is in the Inference tab and installs release v1.10.1 by default. The original Forge is in the Legacy tab. Forge Neo is the actively developed alternative.

Can ComfyUI and Forge share the same models?

Yes. ComfyUI reads the shared folder through extra_model_paths.yaml, and the Forge family through folder links, so an empty models/checkpoints folder in ComfyUI is normal.

Does it support Intel Macs?

No. The official macOS build is for Apple Silicon only.

Related guides and tools

• How to run Stable Diffusion locally
• Stable Diffusion image editors: which are still maintained
• ComfyUI templates
• Checkpoints, LoRAs and VAEs explained
• ComfyUI models and VRAM
• Image generation VRAM calculator

Sources

• LykosAI/StabilityMatrix on GitHub (README, package classes, install and analytics source, releases)
• Stability Matrix docs: installation, data directory, supported packages, installing packages, shared folders and settings
• Lykos AI: End User License Agreement and Privacy Policy
• Upstream repositories via the GitHub API, including reForge, AUTOMATIC1111, Forge and Forge Classic and Neo
• SimpleSDXL issue #97, linked from the package’s advisory

Updated 1 November 2026. Stability Matrix’s source (main branch, commit of 20 October 2026), docs, releases, EULA and privacy policy, and the GitHub status of all 30 packages, were read on that date.