Alternative Installation: Comfy Desktop, macOS, and Linux Manual Installation
The Windows portable package from the previous lesson is only one of the official installation routes. If you use a Mac or Linux, it won't help; if you want a "normal software" experience—graphical installer, automatic updates, multi-instance management—what you need is the official desktop app Comfy Desktop, commonly known in the Chinese-speaking community as the ComfyUI Desktop version. In the words of the official README: the desktop app is "the easiest and best way for new users to use ComfyUI".
This lesson walks through the other two local routes: downloading, installing, and setting up Comfy Desktop (on Windows / macOS / Linux), and the git clone manual installation for macOS and Linux; finally, it explains how to upgrade after installation and whether you can switch routes. If you are already running the portable package, there's no need to reinstall—Step 5 covers how to bring it under desktop app management.
What You'll Be Able to Do After This Lesson
- Install Comfy Desktop (the ComfyUI desktop app) on Windows, macOS (Apple Silicon), or Linux, and complete the first-run setup wizard
- Understand every option in the first-run setup: cloud or local, the three checkboxes, quick installation, and the Chinese mirror toggle for users in mainland China
- Complete a ComfyUI manual installation (macOS and Linux) with git clone + pip, installing the right PyTorch for your GPU
- Upgrade each route: desktop automatic updates, the instance Update tab, manual installation
git pull - Clearly explain how to choose among the four installation routes (portable / desktop / manual / cloud) and how to switch after installing
Your Current Predicament
- I'm a Mac or Linux user, and the Windows portable package from the previous lesson has nothing to do with me; I searched "ComfyUI Mac Installation" and "ComfyUI Linux Installation", and the tutorials that come up all say different things—and they're all several years old
- I want "normal software": double-click the installer, automatic updates, without having to dig through folders for
.batand manually open a browser every time - Manual installation tutorials are full of terms like conda, PyTorch, and CUDA, and I don't know which ones are required; last time I followed one, all the dependencies went into my system Python and conflicted with other software
When to Use This Approach
When you need to:
- Install ComfyUI on a Mac (Apple Silicon) or Linux
- A ComfyUI desktop app with a graphical installer, automatic updates, and snapshot management
- Full control over the Python environment and launch parameters; or to run on a headless Linux server (the desktop app explicitly does not support headless servers)
- Keep multiple ComfyUI versions at the same time and switch per project
And you don't want to:
- Get confused by the "install Python → configure CUDA → install PyTorch" sequence
- Install dependencies into the system Python where they conflict with other software
- Manually re-download and re-extract the archive every time you upgrade
🎒 Preparation Before Starting
- Finish the previous lesson What is ComfyUI: know what nodes, connections, and workflows are, so you won't be lost when you open the interface after installing
- Confirm your system version: the desktop app requires Windows 10 or later / macOS 13 (Ventura) or later / Ubuntu 22.04 or later recommended for Linux
- Confirm your hardware: Mac requires Apple Silicon (M1 or later); memory starting at 8 GB with 16 GB officially recommended; reserve at least 4.85 GB of disk per instance
- If you're taking the manual installation route: install Git, and use the officially recommended Python 3.13
- A cup of water ☕
⚠️ Download Tips for Mainland China
ComfyUI source code and models are mostly hosted on GitHub / Hugging Face, and the official documentation notes that users in mainland China may need to configure a proxy for stable access to GitHub repositories. The desktop app has a built-in Chinese mirror toggle (Step 2), and for manual installation you can switch pip to the Tsinghua mirror (Step 4).
System Requirements: Desktop and Manual Installation Quick Reference
| Item | Comfy Desktop (desktop app) | Manual Installation (git clone) |
|---|---|---|
| Windows | Windows 10 or later, x64 / ARM64 | Also supported (Windows users can also use the portable package from the previous lesson) |
| macOS | macOS 13 (Ventura) and above, Apple Silicon (M1 and later) | Apple Silicon (M1–M4); the official wording is simply "a more recent version of macOS" |
| Linux | Ubuntu 22.04 or newer recommended, x64 (.deb / AppImage) | Any distribution that can run Python 3.13 |
| RAM / disk | 8 GB minimum, 16 GB officially recommended; at least 4.85 GB per standalone installation | Not listed separately by the official docs (the system requirements page only lists platforms, Python versions, and supported GPU types, with no RAM/disk figures) |
| GPU | A discrete GPU (NVIDIA / AMD) is recommended but not required; however, the Windows desktop app only supports NVIDIA CUDA GPUs (Windows users with AMD/Intel GPUs should use the portable package or a manual installation) | NVIDIA / AMD (ROCm) / Intel Arc / Apple Silicon all work; with no discrete GPU you can use --cpu |
| Administrator privileges | Not needed for the Windows installation; required for the first system-level installation on macOS; on Linux, installing .deb requires sudo, AppImage does not | Follow the usual requirements of your system |
Python version only matters for manual installation: the official recommendation is 3.13; 3.14 works but some custom nodes may have issues; 3.12 is the fallback when you run into dependency problems. With the desktop version you don't need to worry about Python—it manages it for you.
Core idea
**Comfy Desktop is not another version of ComfyUI; it is an "instance manager". ** Official definition: a desktop app that lets you install, manage, and launch multiple ComfyUI instances from a single launcher. You don't need to install ComfyUI separately—the Python environment and dependencies are all handled by it; each instance has its own ComfyUI version, custom nodes, and settings, and they can coexist.
The official FAQ puts it plainly: the desktop version does not replace the portable package / manual installation; it is simply the option "recommended for most users"—because it saves effort and comes with one-click updates and snapshot management.
Note a distinction between old and new: "desktop version" in older tutorials online may refer to the legacy Legacy Desktop (a single-instance app). The current Comfy Desktop is a completely rewritten multi-instance manager, and the official docs provide a migration path from the old version.
Manual installation means "setting up the environment yourself". The official docs summarize the steps as four: create a virtual environment (to avoid polluting the system Python) → clone the code repository → install the GPU build of PyTorch and dependencies → launch with python main.py. The desktop version and the portable package hide the first three steps; manual installation means doing them yourself by hand, in exchange for control over every step.
The four routes in one table:
| Route | Who it's for | In one sentence |
|---|---|---|
| Windows portable package (previous lesson) | Windows users who want to get running fastest | Unzip and go; the official README also reminds: the desktop app is more recommended for ordinary users |
| Comfy Desktop desktop version (steps 1–3 of this lesson) | Want a graphical installer, automatic updates, and multi-instance management | Officially called "the easiest and best" for new users |
| Manual installation (step 4 of this lesson) | Want full control over the environment; Linux servers without a graphical interface | Build the official four steps yourself by hand |
| Comfy Cloud | Your local GPU isn't strong enough | Runs in the cloud, nothing installed locally (not covered in this lesson) |
💡 Whichever route you take, you're installing the same ComfyUI
The node interface, workflows, and model directory rules are exactly the same. After installation, the steps for placing models and generating your first image apply site-wide: Where to put models.
Follow along
Step 1: Download and install Comfy Desktop (the ComfyUI desktop version)
Why: The desktop version is the route officially recommended for new users; one installer takes care of ComfyUI, the Python environment, and all dependencies.
Follow the instructions for your platform:
Option A: Windows (.exe installer)
- Open the official download page https://comfy.org/download; it will automatically detect your platform—download the
.exeinstaller - Double-click to run it (NSIS installer) and follow the prompts to the end
- Launch Comfy Desktop from the Start menu or the desktop shortcut
You should see: the Comfy Desktop window opens and enters the full-screen first-time setup (step 2). The official FAQ confirms: installing on Windows does not require administrator privileges.
Option B: macOS (.dmg, Apple Silicon)
- Download the
.dmgfile from https://comfy.org/download - Open the
.dmgand drag Comfy Desktop into the Applications folder - Launch from Launchpad or Spotlight
You should see: on first launch, macOS may show a security warning—go to System Settings → Privacy & Security and click Open Anyway to allow it. This is the normal process documented by the official docs.
Option C: Linux (.deb or AppImage)
Linux doesn't use comfy.org/download; use the general download link given on the official repository page https://dl.todesktop.com/241130tqe9q3y (it auto-detects the platform and offers both .deb and AppImage packages):
# Debian / Ubuntu:在 .deb 文件所在目录执行
sudo apt install ./*.deb
# 或者 AppImage(不需要管理员权限)
chmod +x ./*.AppImage
./*.AppImageYou should see: the Comfy Desktop window opens and enters the full-screen first-time setup.
- Comfy Desktop can now launch on my system
⚠️ If Linux won't start, check two things first
The official docs give two common causes: ① AppImage requires FUSE 2—install it: sudo apt install libfuse2; ② on Ubuntu 24.04+, AppArmor restricts the Electron sandbox—in that case you can add --no-sandbox and retry, but the official docs warn that this flag disables the per-process Electron sandbox, so use it only when the sandbox fails to start. For more troubleshooting, see What to do if installation fails.
Step 2: Complete the first-time setup: cloud or local, Chinese mirror
Why: The first time you open Comfy Desktop, it enters a full-screen setup and won't proceed until you answer all the questions; if you close the app midway, this same screen appears the next time you open it.
Let's go through each setting item:
| Setting | How to handle it | Notes |
|---|---|---|
| Cloud or Local | Choose Local to run locally | Local installs ComfyUI on your machine: you control the version, custom nodes, and launch parameters, and it can run offline; Cloud runs on Comfy Cloud's GPUs with nothing installed on your machine |
| EULA and Terms of Service | Must be checked | You cannot continue without checking it |
| Anonymous usage data | Optional, checked by default | Official statement: no workflow content, prompts, generated media, or personal information is collected; you can change this later in settings |
| Beta features | Optional, checked by default | Only available when data sharing is enabled |
| Express Install | Optional, unchecked by default | If checked, it skips the installation wizard and installs a Standalone instance directly with recommended settings; if unchecked, you first see ComfyUI version, installation location, and launch options before anything is written to disk. The detected hardware is shown below the checkbox, and you can cancel if the detection is wrong |
Two details: when no card is selected, Continue is grayed out—the official stance is that you should think it through before choosing; when your local GPU isn't powerful enough, the Cloud card gets a "Recommended for your hardware" badge. If you installed the old Legacy Desktop, there is also an extra Migrate existing instance checkbox (checked by default) that migrates your workflows, models, and settings.
💡 Users in mainland China: enable the Chinese mirrors
When the app interface language is Chinese, the first-time setup will ask one extra question after Continue: Enable Chinese Mirrors? Choose Enable Mirrors, and the Git repository will be cloned from gitcode.com while PyPI packages go through the Alibaba Cloud and Tencent mirrors (replacing github.com and pypi.org)—officially stated to be faster in mainland China. Choosing Not now to keep the defaults is also fine, and you can change it anytime later in settings (the setting is called Use Chinese Mirrors (Git & PyPI)).
After clicking Continue:
You should see: a brief Preparing your install…, then the Local route enters the installation progress (the installation wizard runs first if Express Install is unchecked), and the completed instance appears on the dashboard.
- My first instance appeared on the dashboard
Step 3: Launch an instance and get to know Comfy Desktop's daily operations
Why: The daily routine of the desktop version is "managing instances on the dashboard"—starting, stopping, and upgrading all happen on the instance cards.
Click a card to launch an installed instance; ComfyUI runs in a standalone window (official FAQ: multiple instances can run at the same time, each with its own independent Python environment). Three common operations:
| Operation | How to do it |
|---|---|
| Stop an instance (keep the app) | The ⋮ menu on the instance card → Stop → confirm. This stops the ComfyUI Python backend; the desktop app stays open and can be restarted anytime |
| Quit the entire app | The ☰ menu in the top-left corner → Quit Desktop (or use your system's usual quit method) |
| Add an instance | On the dashboard, New Instance: Standalone (fresh install with its own Python), Add Existing Instance (bring a git clone or the portable package from the previous lesson under management), Cloud (connect to the cloud) |
Remember one rule about model directories first: in the desktop version, instances share the model library by default, and the shared folder is configured in Desktop Settings → Storage. For a detailed explanation of directories and sharing models across multiple installations, see Where to put models.
You should see: after clicking the card, the instance enters the running state, ComfyUI opens in a standalone window; the instance card's ⋮ menu contains Stop, and after clicking it the app remains while the instance stops.
- I can start and stop an instance, and I know where New Instance is
Step 4: Manual installation of ComfyUI: git clone plus pip (macOS and Linux)
Why: If you want full control over the environment, or to run on a Linux server without a graphical interface (the desktop FAQ explicitly says headless / server mode is not supported), follow the official four steps: create a virtual environment → clone the repository → install dependencies → launch.
Manual installation on Windows is also officially supported, but with one prerequisite: install the Microsoft Visual C++ Redistributable first. However, Windows users already have the portable package and the desktop version, so this route is generally unnecessary.
💡 There is also an official command-line tool, comfy-cli
The official documentation also offers comfy-cli: pip install comfy-cli, then comfy install—installing ComfyUI via the command line with dependencies managed for you. This lesson follows the conda + git workflow recommended by the official documentation; comfy-cli is left for you to explore on your own if you prefer the command line.
Method A: Manual installation on macOS (Apple Silicon)
Prerequisites: install Git and Miniconda (officially used to create an isolated Python environment).
# 1. 创建并进入虚拟环境(避免污染系统 Python)
conda create -n comfyenv
conda activate comfyenv
# 2. 克隆 ComfyUI 官方仓库
git clone https://github.com/Comfy-Org/ComfyUI.git
# 3. 装 PyTorch nightly(官方指定 Mac 走 nightly 版)
conda install pytorch-nightly::pytorch torchvision torchaudio -c pytorch-nightly
# 4. 装 ComfyUI 依赖并启动
cd ComfyUI
pip install -r requirements.txt
python main.pyYou should see: the terminal finally outputs To see the GUI go to: http://127.0.0.1:8188, and opening this address in a browser shows the node interface. To confirm that Apple Silicon GPU acceleration (MPS) is working, run one more official verification command:
python -c "import torch; print(torch.backends.mps.is_available())"An output of True means Metal acceleration is available. Official note: on Mac, ComfyUI uses PyTorch's MPS (Metal Performance Shaders) backend, not Apple's MLX.
Method B: Manual installation on Linux (choose PyTorch based on your GPU)
Steps 1, 2, and 4 are exactly the same as on macOS (create the environment with conda, clone the repository, install dependencies and launch); only step 3 needs to be replaced with the PyTorch matching your GPU:
# NVIDIA(CUDA 13.0 稳定版)
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
# AMD(ROCm 7.2 稳定版)
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2
# Intel Arc(XPU 后端;官方提醒:不要装 CUDA 版 wheel)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpuThe official documentation also offers a nightly version of PyTorch (which may include performance improvements, but is labeled "higher risk" by the officials themselves); the stable version above is fine for beginners. Then:
cd ComfyUI
pip install -r requirements.txt
python main.pyYou should see: the same To see the GUI go to: http://127.0.0.1:8188 at the end, and opening it in a browser shows the node interface. A manually launched ComfyUI is stopped by pressing Ctrl+C or simply closing the terminal window (verbatim from the official FAQ).
💡 Users in mainland China: switch pip to the Tsinghua mirror
The solution given in the official documentation: pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple. If git clone is slow or fails, the official documentation mentions that users in mainland China may need to configure a proxy to access the GitHub repository reliably.
- The terminal shows
To see the GUI go to: http://127.0.0.1:8188, and the node interface opens in a browser
Step 5: Upgrading and switching routes
Why: ComfyUI iterates quickly; installation is a one-time thing, but upgrading is long-term.
Comparison of upgrade methods for each route (all from the official documentation):
| Route | Upgrade method |
|---|---|
| Comfy Desktop app | checks for updates automatically, cannot be turned off (once shortly after launch, then every 10 minutes). When the Desktop Update Ready button appears in the title bar, click it to restart and install; you can also go to Desktop Settings → Updates to manually Check for updates |
| In the desktop version, the ComfyUI instance | and app updates are two different things: use the instance's Update tab and choose the Stable or Latest on GitHub channel |
desktop version installed via Linux .deb | in-app updates do not apply—re-download using the universal download link, then sudo apt install ./*.deb |
| Manual installation | first cd <安装路径>/ComfyUI, then git pull, then pip install -r requirements.txt. ⚠️ Confirm the virtual environment is activated first, otherwise dependencies will be installed into the system Python |
Switching routes: Portable package users don't have to start over—the desktop version's New Instance → Add Existing Instance officially supports bringing an existing portable package or git clone under management; models can also be shared via extra_model_paths.yaml between multiple installations, see Where to put models for how.
What you should see (taking desktop auto-update as an example): the title bar first shows Downloading update…, then changes to Desktop Update Ready when ready; click it to restart and the install completes—this is exactly the default auto-update flow described officially.
- I know how to upgrade my own route
Checkpoint ✅
- Comfy Desktop can launch, and my first instance appears on the dashboard
- I understood every option in the first-time setup; the Chinese mirror is enabled or I know where to change it
- (Manual route) The terminal showed
To see the GUI go to: http://127.0.0.1:8188, and the browser can open the node interface - I know how to upgrade my route: desktop auto-update, the instance Update tab, manual
git pull - I can say which of the four routes suits me, and how to switch after installing
Pitfall reminders
❌ macOS: a security warning pops up when first opening Comfy Desktop and the app won't start
A security warning on first launch is a normal phenomenon explicitly stated in the official documentation, not a broken installation. Fix: System Settings → Privacy & Security → Open Anyway. If it still won't start, the official FAQ suggests going to the download page to confirm you have the latest version.
❌ Linux: after installing the AppImage or .deb, the app won't open
Two causes listed officially: missing FUSE 2 (required by AppImage, sudo apt install libfuse2); or the distribution blocking the Electron sandbox (AppArmor on Ubuntu 24.04+ restricts unprivileged user namespaces), in which case add --no-sandbox and retry—note the official warning: this flag disables the Electron sandbox for every process, so use it only when the sandbox fails to start.
❌ Manual installation: running pip install without activating the virtual environment
Symptoms: dependencies get installed into the system Python and conflict with other software; the next upgrade pip install -r requirements.txt will also install to the wrong place. The official warning, paraphrased: before installing dependencies, you must confirm the current Python environment is a ComfyUI virtual environment. Correct order:
conda activate comfyenv # 先激活
pip install -r requirements.txt # 再装依赖❌ Manual installation fails on startup with "Torch not compiled with CUDA enabled"
The official README's troubleshooting item: first uninstall torch completely, then reinstall using the command for your GPU.
pip uninstall torch
# NVIDIA 示例:重装 cu130 稳定版
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130Advanced tips (optional)
🔧 The desktop version creates snapshots automatically: According to the official documentation, snapshots are created automatically on key events—every launch / restart, before and after updates, and after a restore; the contents are the ComfyUI version, custom node versions, and all pip packages (not including models or input/output files). In the instance's Snapshots tab you can browse, restore, export, and share. When an update fails, the official advice is to manually create a snapshot first, then retry.
🔧 Windows desktop version repeatedly exits on its own after updating: Officially confirmed that 1.0.46 and earlier versions have a bug—if the update download hasn't finished, the app flashes the update screen and closes itself, and this happens on every launch, too fast for self-repair. Fix: manually download the latest Windows installer (https://dl.todesktop.com/241130tqe9q3y/windows or the download page) and install directly, no need to uninstall the old version first; instances, models, and settings are all preserved. Fixed since 1.0.47.
🔧 Want to completely uninstall the desktop version: Uninstalling the app only removes the launcher itself; instances, shared models, settings, and logs all remain on disk. The official documentation lists the leftover directories per platform: Windows in %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Installs, %APPDATA%\Comfy Desktop, etc.; macOS in ~/ComfyUI-Installs, ~/Library/Application Support/Comfy Desktop, etc.; Linux in ~/.config/comfyui-desktop-2 etc. To clean up completely, delete them one by one following the official page for your platform.
🔧 There is no such thing as an "ComfyUI Docker official image": The official FAQ explicitly states no official Docker image is provided; community images on Docker Hub are not officially maintained and unsupported officially—use at your own risk. For server deployment, take the manual installation route.
FAQ
ComfyUI download and installation tutorial? Where can I download the desktop version?
For the desktop version (Comfy Desktop), go to the official download page https://comfy.org/download, which automatically detects your platform: on Windows you get the .exe installer, on macOS .dmg, while Linux does not use this page — instead, use the general download link given on the official repository page https://dl.todesktop.com/241130tqe9q3y (offering both .deb and AppImage packages). After downloading, double-click the installer or drag the app into Applications; see "Step 1" on this page for detailed instructions.
ComfyUI How do I download and install it? What are the manual installation steps?
The official guide summarizes manual installation in four steps: create a virtual environment (to avoid polluting the system Python) → clone the code repository → install the GPU version of PyTorch and dependencies → launch with python main.py. For the full commands on macOS and Linux, see "Step 4" on this page; the sign of a successful launch is the terminal output To see the GUI go to: http://127.0.0.1:8188.
ComfyUI What are the system requirements?
The desktop version requires Windows 10 or later / macOS 13 (Ventura) or later (and requires Apple Silicon, M1 and later) / Linux, where Ubuntu 22.04 or later is recommended; memory starts at 8 GB with an official recommendation of 16 GB, and you should reserve at least 4.85 GB of disk space per instance. As for graphics cards, a discrete GPU is recommended but not required; however, the Windows desktop app only supports NVIDIA CUDA GPUs. With manual installation, NVIDIA / AMD / Intel Arc / Apple Silicon all work, and you can use --cpu if you have no discrete GPU.
ComfyUI Can it be installed on a Mac M4?
Yes. The desktop version requires Apple Silicon (M1 and later) on macOS 13 (Ventura) or above; the M4 is within the supported range; the manual installation route also supports Apple Silicon (M1–M4). On Mac, ComfyUI uses PyTorch's MPS (Metal Performance Shaders) backend; for manual installation, the official guidance specifies installing the PyTorch nightly version, and after installation you can verify it with python -c "import torch; print(torch.backends.mps.is_available())" — if the output is True, Metal acceleration is available.
ComfyUI: How is it used and what is the experience like? How does it differ from the portable package?
Comfy Desktop is not another version of ComfyUI, but rather an "instance manager": you don't need to install ComfyUI separately — it manages the Python environment and dependencies for you, and also comes with a graphical installer, automatic updates, and snapshot management. Officially, it is described as the "easiest and best" way for new users to use ComfyUI. Whichever route you take, you end up installing the same ComfyUI, with the node interface, workflows, and model directory rules all exactly the same.
Lesson summary
✅ ComfyUI Desktop version (Comfy Desktop): you can now install and use it—download the installer → full-screen first-time setup (Cloud / Local, three checkboxes, Chinese mirror) → launch the instance from the dashboard.
✅ macOS (Apple Silicon) and Linux users have two paths: the desktop version (.dmg drag into Applications / .deb / AppImage), or git clone + pip manual installation—the official four steps: create a conda environment, clone the repository, install PyTorch for your GPU, python main.py.
✅ You've sorted out the upgrade routes: the desktop version checks for updates automatically (can't be turned off), instances use the Update tab, manual installation uses git pull; a portable package can later be brought under desktop management via Add Existing Instance, no need to start over.
| Command / URL | Purpose |
|---|---|
https://comfy.org/download | Comfy Desktop official download page (auto-detects platform) |
https://dl.todesktop.com/241130tqe9q3y | Linux universal download link (.deb / AppImage) |
git clone https://github.com/Comfy-Org/ComfyUI.git | Manual installation: clone the ComfyUI source code |
conda create -n comfyenv / conda activate comfyenv | Create / activate the virtual environment |
pip install -r requirements.txt | Install ComfyUI dependencies |
python main.py | Launch ComfyUI (manual route) |
git pull | Upgrading ComfyUI source code (manual route) |
Next Lesson Preview
Finishing the installation doesn't mean everything will go smoothly: What to Do If Installation Fails: Installation and Startup Troubleshooting.
You will learn:
- The order of troubleshooting when ComfyUI fails to start or a black window flashes and disappears
- What to do if
127.0.0.1:8188won't open or the interface shows a black screen - How to handle startup errors related to graphics drivers
- Where to find the log files when the desktop version fails to start
If everything is already running smoothly, you can go straight to Where to Put Models to place your first model and generate images sooner; for the full image-generation workflow, see First Image Generation.