How to install ComfyUI on Windows and Linux, step by step

The easiest way to install ComfyUI on your own computer is the Comfy Desktop app: it exists for Windows and macOS, and the installer is on comfy.org/download. On Linux you install ComfyUI from source with git clone, PyTorch for CUDA and python main.py, and the interface opens at http://127.0.0.1:8188. If your video memory falls short (FLUX.1 in fp8 needs at least 13.3 GB), run ComfyUI on a rented GPU in 2–5 minutes.

Which way of installing should you pick?

There are three official ways, and each suits different systems. The ComfyUI authors recommend Desktop to newcomers and say the Windows portable build is not meant for regular users: with it you track updates yourself.

How much video memory does ComfyUI need?

ComfyUI itself weighs almost nothing; the video memory goes to the model. The table is our estimate for seven setups. “Minimum” means ComfyUI swaps the text encoder out before sampling; “comfortable” means everything sits in video memory at once. SDXL at the minimum runs on a 8 GB card, FLUX.1 in fp8 on a 16 GB one.

Mind the weight format. FLUX.1 in bf16 needs 24.3 GB even at the minimum, which is why it runs in fp8 on a home card. Video asks for more: Wan 2.2 A14B, with its two experts, takes 40.6 GB comfortably. The other models are compared in the image models VRAM guide, and you can test your own combination in the calculator.

How to install ComfyUI on Windows

For a first install, take Comfy Desktop. Download the installer from comfy.org/download, run it and follow the prompts. On first launch the app asks whether to work on Comfy's cloud or locally on this machine: choose local. Installs live in %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Installs, the shared model library in %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared, and the log in %APPDATA%\Comfy Desktop\logs.

If you want the portable build for NVIDIA, follow these steps.

1. 1

   Download the archive

   ComfyUI_windows_portable_nvidia.7z is on the releases page of the comfyanonymous/ComfyUI repository on GitHub, and the README has a direct link. You do not need any other download site.

2. 2

   Extract it

   7-Zip or Windows Explorer will do. If the archive will not extract, right-click the file, open Properties and press Unblock.

3. 3

   Find the right launcher

   The folder holds ComfyUI, python_embeded, update and the launchers. For an NVIDIA card you want run_nvidia_gpu.bat; run_cpu.bat starts ComfyUI without a GPU, slowly.

4. 4

   Run run_nvidia_gpu.bat

   A console window opens. Wait for the line “To see the GUI go to: http://127.0.0.1:8188”. The browser usually opens by itself; if it does not, type the address in.

5. 5

   Put a model in its folder and refresh the page

   Checkpoint files (safetensors) go into ComfyUI\models\checkpoints. Larger models come in several files, each with its own subfolder in ComfyUI\models.

Keep the console window open: closing it stops ComfyUI. To update, the update folder has scripts: update_comfyui.bat fetches the latest commit and update_comfyui_stable.bat the latest stable release.

How to install ComfyUI on a Mac

A Mac needs an Apple Silicon chip (M1 or later) and macOS 13 or later. Open the downloaded .dmg file, drag Comfy Desktop into the Applications folder and launch it from Launchpad or Spotlight. If macOS shows a security warning, go to System Settings, Privacy & Security, and press Open Anyway. The documentation recommends at least 4.85 GB for the installation.

How to install ComfyUI on Linux

There is no Desktop app for Linux, so you install ComfyUI by hand. You need Git and Python: the authors recommend 3.13, or 3.12 if custom-node dependencies give trouble (3.14 works, but some nodes may misbehave). You can make the environment with venv or conda; the documentation uses conda. For an NVIDIA card install PyTorch from the CUDA 13.0 index (cu130), which cards of the 20 series and newer require.

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python3 -m venv venv
source venv/bin/activate
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt
python main.py

When the terminal prints the address line, open http://127.0.0.1:8188 in a browser. Press Ctrl+C to stop ComfyUI. On AMD cards under Linux, install PyTorch from the ROCm 7.2 index, on Intel from the xpu index. They do not need the CUDA packages:

# AMD (Linux)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2

# Intel
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

There is also a shortcut through the comfy-cli command line: pip install comfy-cli, then comfy install. Models go into the same folders as on Windows: checkpoints in models/checkpoints, the VAE in models/vae.

How to add ComfyUI-Manager

ComfyUI-Manager installs custom nodes and updates them from the interface. In the ComfyUI repository it is switched on with a flag. From the ComfyUI folder run two commands:

pip install -r manager_requirements.txt
python main.py --enable-manager

The ComfyUI-Manager guide covers the details, working with nodes and rolling back a bad install.

How to reinstall or update ComfyUI

Save your models first: they weigh far more than the program. In the portable build they sit in ComfyUI\models, in Desktop in the shared ComfyUI-Shared folder. Copy them, extract a fresh archive into a new folder or install Desktop again, then put the models back.

To avoid copying gigabytes, point to the old folder in extra_model_paths.yaml: rename extra_model_paths.yaml.example in the ComfyUI root, edit the paths and restart. If something breaks after an update, start with the log. In Desktop that is app.log.

What if your graphics card is too weak?

If your card is smaller than the “Minimum” column, do not torture it. FLUX.1 in fp8 needs a 16 GB card, and 24 GB to work comfortably. You need such a card for hours, not for good, so renting is simpler. ComfyUI is already in the template on the ComfyUI page, and the full cloud route is in the server install guide.

1. 1

   Pick a card from the table

   A 24 GB card such as the RTX 4090 is enough for SDXL and FLUX.1 in fp8. For bf16 and video take a bigger one.

2. 2

   Start a server from the ComfyUI template

   Open the ComfyUI page and press Rent. The server starts in 2–5 minutes.

3. 3

   Open the interface

   The ComfyUI button on the server card opens the same node editor you have on your own computer.

4. 4

   Load a model and a workflow

   It is easier to download files straight onto the server: the server install guide shows how.

5. 5

   Stop the server when you are done

   Billing is per minute. A stopped server costs nothing for the GPU, only its disk.

Choosing a model is covered by the FLUX guide and the Stable Diffusion guide. To see in advance which card your own combination needs, use the VRAM calculator.

What can go wrong

• ComfyUI will not open at 127.0.0.1:8188. Check that the console window is open and shows the address line. Another process may hold the port: start ComfyUI with --port 8189 and open the new address. 127.0.0.1 works only on the same computer; from another device you need the --listen flag.
• The error “Torch not compiled with CUDA enabled”. The environment has a PyTorch without CUDA. Remove it with pip uninstall torch and install it again from the cu130 index in the block above.
• The portable build will not start. Update the NVIDIA driver. Check that you took the archive for your card series: the cu126 build is not for the 20 series and newer.
• Out of video memory. Lower the resolution and take the model in fp8 or a quantised version. If that does not help, the card is short on memory by the table above.
• Unprotected access from the local network. The --listen flag opens the interface to everyone who can see your computer. Use it only on a network you trust.

How we counted

Commands and file names come from the README of comfyanonymous/ComfyUI and from docs.comfy.org; the default address and port come from comfy/cli_args.py. We made up no system requirements: where a page gives no number (disk, RAM), none appears here. Model memory is computed from the sizes of the published files plus the text encoder, the VAE and working memory: the minimum is the larger of “encoder plus 0.5 GB” and “network plus VAE plus working memory”. The working memory is our estimate, not measured on GPUs. 1 GB here is 1024³ bytes, and a card's usable memory is 97.5% of its nominal size. Data as of 2026-10-19; the other models are in the calculator.