How to Install ComfyUI on Windows, Mac and Linux (2026): Desktop, Portable and Manual

ComfyUI is the most powerful free tool for running AI image and video models on your own PC. The catch is that it now comes in three different flavours, and picking the wrong one is the number one reason beginners get stuck.

In this guide I walk you through all three install methods on Windows (plus Mac and Linux), show you how to make your first image, and fix the errors people hit most often. Everything here matches the official ComfyUI docs as of November 2026.

Desktop vs Portable vs Manual: Which to Pick

All three methods run the exact same ComfyUI. The difference is how much of the setup is done for you, and how easy it is to tinker later.

What You Need Before You Start

ComfyUI will run on almost any modern PC, but your GPU decides which models you can use and how fast they run.

Tip: update your NVIDIA driver before you install. The CUDA 13.0 builds need a recent driver (the 580 series or newer), and an old driver is the most common reason the portable build closes straight away.

Not sure how much VRAM you have? My guide on how to check GPU VRAM shows you in 30 seconds. For which models fit your card, see how much VRAM you need for AI.

Method 1: Install Comfy Desktop (Easiest)

Comfy Desktop is now a launcher that installs and manages one or more ComfyUI instances. It handles Python, PyTorch and dependencies for you, and models are shared across all instances by default.

1. Download the installer. Go to comfy.org/download. The page detects your platform and gives you the right file (a .exe on Windows).
2. Run the installer. Double-click the .exe and follow the prompts. Then open Comfy Desktop from the Start Menu or desktop shortcut.
3. Choose Local or Cloud. The first screen asks “How do you want to run Comfy?”. Pick Local to run on your own GPU, offline. Cloud runs on Comfy’s paid GPUs instead.
4. Accept the terms. You must accept the EULA and Terms of Service. Anonymous usage data sharing is on by default; you can switch it off here.
5. Use Express Install (optional). Turn on Express Install to create a Standalone instance with recommended settings for your detected hardware. Leave it off if you want to choose settings in the wizard.
6. Migrate an old install (optional). If you used the older Desktop app, “Migrate existing instance” is on by default. It carries over custom nodes, workflows, models and settings.
7. Launch your instance. When the install finishes, start the instance from the dashboard. The ComfyUI canvas opens inside the app.

Note: if your GPU is weak, the Cloud card shows a “Recommended for your hardware” badge. You can still choose Local and use small, fast models.

Where Comfy Desktop keeps your files on Windows

Older installs used %USERPROFILE%\ComfyUI-Installs and %USERPROFILE%\ComfyUI-Shared instead. Paste any of these paths into the File Explorer address bar to open them.

Method 2: Windows Portable Build

The portable build is a single folder with its own embedded Python. Nothing is installed system-wide, so you can copy it to another drive or keep two versions side by side.

Pick the right download

All of these are on the Releases page of the Comfy-Org/ComfyUI GitHub repository, under the latest release.

Install and run it

1. Install 7-Zip. Get it from 7-zip.org. The archives are .7z files and need a proper extractor.
2. Extract to a short path. Right-click the archive, choose 7-Zip, then Extract Here. Use something like D:\AI\ rather than your Desktop, to avoid long-path problems.
3. Open the folder. You will see ComfyUI, python_embeded, update, README_VERY_IMPORTANT.txt and the launcher .bat files.
4. Double-click your launcher. Use run_nvidia_gpu.bat, run_amd_gpu.bat or run_intel_gpu.bat. Use run_cpu.bat only if you have no GPU.
5. Wait for the address. When the console shows “To see the GUI go to: http://127.0.0.1:8188”, your browser should open. If not, type that address yourself.

Warning: keep the black console window open. Closing it stops ComfyUI, and any running generation is lost.

Installing extra Python packages in portable

Always use the embedded Python, never your system Python. From inside the portable folder, run:

.\python_embeded\python.exe -m pip install package-name

To add a launch flag, edit your .bat file. The line looks like this:

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --your-flag

The full list of flags is in my ComfyUI command line arguments guide.

Method 3: Manual Install With Git

The manual route gives you full control and is the standard choice on Linux. It takes about 10 minutes plus the PyTorch download (around 2 to 3 GB).

1. Install Python 3.13 and Git. On Windows, tick “Add Python to PATH” in the Python installer.
2. Clone ComfyUI. Open a terminal in the folder you want and run git clone https://github.com/Comfy-Org/ComfyUI.git.
3. Create a virtual environment. Run cd ComfyUI, then python -m venv venv. Activate it with venv\Scripts\activate on Windows or source venv/bin/activate on Linux and Mac. Conda (conda create -n comfyenv) also works.
4. Install PyTorch for your GPU. Use the command from the table below. This step matters most.
5. Install ComfyUI’s requirements. Run pip install -r requirements.txt.
6. Start ComfyUI. Run python main.py and open http://127.0.0.1:8188.

PyTorch install command for each GPU

Tip: check that PyTorch sees your GPU before going further. Run python -c "import torch; print(torch.cuda.is_available())". It should print True on NVIDIA and AMD ROCm.

AMD owners can find card-by-card details in my AMD GPU for AI guide. RTX 50 owners should also read ComfyUI on RTX 50 series.

The comfy-cli shortcut

If you prefer one command, the official comfy-cli tool clones ComfyUI and installs dependencies for you:

pip install comfy-cli
comfy install
comfy launch

Installing on Mac and Linux

• Mac (Apple Silicon): Comfy Desktop supports macOS 13 and newer on M-series chips. The manual route also works, using PyTorch’s MPS backend. Intel Macs are not supported by Desktop.
• Linux: Comfy Desktop ships as a .deb and an AppImage, but most Linux users go manual so they can pick the exact PyTorch build.
• Remote or headless: start with python main.py --listen to reach ComfyUI from another device on your network at http://your-pc-ip:8188.

No GPU at all? Renting one for under a dollar an hour is often easier. My ComfyUI on RunPod guide covers the official template.

Make Your First Image

A fresh ComfyUI has no models, so the canvas cannot make anything yet. The quickest fix is a built-in template, which tells you exactly which files to download.

1. Open the templates. Click the Workflow menu (or the Templates button in the sidebar) and browse the image templates.
2. Pick a beginner model. Choose a small, fast text-to-image template that matches your VRAM. On 8 GB cards, distilled models like FLUX 3 Image Klein are a good start.
3. Download the missing models. ComfyUI shows a dialog listing the files the template needs. Download them and they go into the right ComfyUI/models/ sub-folders.
4. Write a prompt. Type your description in the positive prompt box, for example “a cosy cabin in a snowy forest at dusk, warm window light”.
5. Click Run. Press the Run button (or Ctrl+Enter). Your image appears in the Save Image node and in ComfyUI/output/.

Want to understand the settings on the sampler node? My guide to ComfyUI samplers and schedulers explains each one in plain words.

Where to Put Models

Every model type has its own folder under ComfyUI/models/. Put a file in the wrong folder and its loader node will not list it.

For every other folder (GGUF, CLIP vision, IP-Adapter and more), use my ComfyUI model folder cheat sheet.

Share models between installs

Already have models from another install or app? Do not copy them. In the ComfyUI folder, copy extra_model_paths.yaml.example, rename it to extra_model_paths.yaml, and point it at your existing folder:

my_models:
    base_path: D:/AI/models
    checkpoints: checkpoints/
    loras: loras/
    vae: vae/
    diffusion_models: diffusion_models/
    text_encoders: text_encoders/

Restart ComfyUI afterwards. In Comfy Desktop, add the folder in Desktop Settings under Storage instead, or edit %APPDATA%\ComfyUI\extra_models_config.yaml after making a backup.

How to Update ComfyUI

Note: ComfyUI ships a new stable version roughly every two weeks. Only run the portable update_comfyui_and_python_dependencies.bat when something is broken, because it reinstalls packages.

What to Install Next

• ComfyUI Manager: lets you install custom nodes and fix “missing node” errors. It is built in now; see how to install ComfyUI Manager.
• Must-have nodes: rgthree, VideoHelperSuite, KJNodes and more in my custom nodes list.
• Speed boost on NVIDIA: SageAttention on Windows can speed up video models a lot.
• Low VRAM: ComfyUI-GGUF lets 8 to 12 GB cards run big models.
• Video: start with Wan 2.2 on 8 to 16 GB.

Troubleshooting Common Install Errors

The portable console opens and closes instantly

This is nearly always an old NVIDIA driver. Install the latest Game Ready or Studio driver and try again. If it still closes, open a Command Prompt in the folder and run the .bat from there so you can read the error.

“Torch not compiled with CUDA enabled”

You have a CPU-only PyTorch. In your venv run pip uninstall torch torchvision torchaudio, then reinstall with the cu130 command from the table above.

“Python is not recognized”

Python is not on your PATH. Reinstall Python and tick “Add Python to PATH”, or use the portable build, which does not need system Python.

My model does not show in the loader

• Check it is in the right folder (diffusion models are not checkpoints).
• Press R in the canvas to refresh the model lists, or restart ComfyUI.
• Check the file finished downloading and is not a .part or .crdownload file.

Out of memory on my first image

The model is too big for your VRAM. Pick a smaller or GGUF version, lower the resolution, or try the --lowvram flag. My CUDA out of memory fix lists every option.

The browser says “Reconnecting”

The ComfyUI server crashed or ran out of RAM. Look at the console for the last error. My Reconnecting fix guide covers nine causes, and 22 common ComfyUI errors covers the rest.

Comfy Desktop install fails

Open %APPDATA%\Comfy Desktop\logs\app.log and scroll to the end. Failed downloads usually come from antivirus or a VPN blocking PyPI; pause them and retry the install.

How to Uninstall ComfyUI Cleanly

• Desktop: Settings, Apps, Installed apps, Comfy Desktop, Uninstall. This removes only the launcher; delete the folders in the Desktop table above to remove models and outputs too.
• Portable: delete the extracted folder. Nothing else was installed.
• Manual: delete the ComfyUI folder (the venv lives inside it). Optionally uninstall Python and Git.

Warning: deleting the shared models folder removes every model you downloaded. Move it to another drive first if you plan to reinstall.

My Take

For most people, Comfy Desktop is now the right answer. It removes the Python and CUDA headaches that used to scare beginners away, and shared models mean you can try a second instance without doubling your disk use.

I still keep a portable copy for testing risky custom nodes, because a broken folder is easy to throw away. On Linux, manual install with a venv is still the cleanest option.

FAQ

Is ComfyUI free?

Yes. ComfyUI is free and open source, and running it locally costs nothing. Only the optional Comfy Cloud and paid API nodes cost money.

Which ComfyUI version should I download for an RTX 5080 or 5090?

Use Comfy Desktop or ComfyUI_windows_portable_nvidia.7z, which ships PyTorch with CUDA 13.0. Do not use the cu126 build; it is meant for GTX 10 series and older cards.

Can I run ComfyUI without a GPU?

Yes, with run_cpu.bat or the --cpu flag, but a single image can take many minutes. A cloud GPU is a better option if you have no graphics card.

Can I have Desktop and portable installed at the same time?

Yes. They are separate. You can even add your portable folder to Desktop with “Add Existing Instance” and share one models folder with extra_model_paths.yaml.

What port does ComfyUI use?

Port 8188 by default, at http://127.0.0.1:8188. Change it with --port 8189 if another program uses 8188.

Does ComfyUI work on AMD GPUs on Windows?

Yes. Use the AMD portable build or AMD’s ROCm 10.0 PyTorch on Windows 11. RDNA 2, 3, 3.5 and 4 cards are supported.

How big is a ComfyUI install?

About 5 GB per install before models. Models add 2 to 30 GB each, so plan for at least 50 GB of free space.

Sources

• ComfyUI docs: Comfy Desktop for Windows
• ComfyUI docs: Comfy Desktop overview and first launch
• ComfyUI docs: ComfyUI Portable for Windows
• ComfyUI docs: Manual install
• ComfyUI docs: System requirements
• Comfy-Org/ComfyUI on GitHub