How to install and set up ComfyUI in 2026

Installing ComfyUI is easy now. The confusing part is that there are four ways to do it, every tutorial picks a different one, and they behave differently afterwards. So this guide starts by helping you choose, then does each route properly, then covers the part most guides skip: what to do in the first twenty minutes after it opens.

Everything here was done on Windows 11 with an RTX 4090, on the Desktop app and the Windows portable build, both running ComfyUI 0.36.0. Where you see a number, like a file size, a folder or a download time, it was read off those installs.

Which install should you pick?

If you're not sure, install the Desktop app. It's the route the ComfyUI README recommends for most people, it manages its own Python, and nothing stops you adding a portable copy later. The two can even share one models folder, which is covered further down.

What you need

The official system requirements list supported platforms but don't give a minimum amount of VRAM, because the answer depends entirely on which model you run. Here's the practical version.

Route A: the Desktop app

The Desktop app runs on Windows and on Apple Silicon Macs (M1 or later, with macOS 13 or newer). There's no Linux version, so on Linux go to Route C. The steps and screenshots below are from Windows.

1. Download it

   Go to comfy.org/download and click Download Desktop. It's free and open source. The cloud product on the same page is the paid one.

   

2. Run the installer

   A standard Windows installer. It checks for the Microsoft Visual C++ Redistributable and installs it if it's missing. Let it: that component is behind a lot of the c10.dll errors people hit later.

   

3. Choose Local

   On first launch it asks how you want to run Comfy. Cloud rents GPUs; Local uses your own. Pick Local, tick the agreement and click Continue. As the screen says, you can switch later.

   

4. Check the new instance

   This screen sets up the actual ComfyUI environment. The defaults are right for most people: Standalone, the Stable release, and the recommended version. What matters is Hardware. Pick the one that matches your graphics card; each line tells you the ComfyUI version, the Python version and how big the download is.

   

5. Skip the starter workflow (probably)

   Next it offers to set you up with a ready-made workflow, and downloads that workflow's models during setup. Look at the sizes before you choose: the recommended video workflow is about 53 GB, and some of the options run on paid credits. Skip & Install gets you a working ComfyUI now, and you can pick a smaller model from the templates afterwards.

   

6. Let it download and unpack

   It now downloads the environment (the 2,256 MB NVIDIA build) and unpacks it. At about 4.5 MB/s that took me roughly twelve minutes from clicking Continue to seeing the canvas. It only happens once, and ComfyUI opens with the template browser when it's done.

Once it's installed, this is where everything lives:

the ComfyUI install itself
%LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Installs

models and outputs, shared between installs
%LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared

app settings and logs
%APPDATA%\Comfy Desktop

I haven't run the Mac version for this guide, so this part is from the official Mac instructions: the download is a .dmg, you drag Comfy Desktop into Applications, and it keeps its installs in ~/ComfyUI-Installs and your models and outputs in ~/ComfyUI-Shared.

Route B: the Windows portable build

One folder, with its own Python inside, and nothing registered with Windows. It's my choice for a second install, because you can copy the folder to another drive and it keeps working.

1. Download the archive

   From the releases page on GitHub, or directly:

   Direct link, NVIDIA buildCopy

   https://github.com/Comfy-Org/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z

   It's about 1.9 GB. The standard build comes with Python 3.13 and PyTorch for CUDA 13.0, and supports RTX 20-series cards and newer. There are separate builds for AMD and Intel, and a cu126 build with Python 3.12 for GTX 10-series and older cards, which the README warns not to use on anything newer.

2. Extract it, and give it a few minutes

   Right-click the file and choose Extract All, or use 7-Zip. It unpacks to 62,899 files and 3.87 GB, which took about six minutes here, so don't worry about the file count.

   

   If Windows blocks the .bat Files downloaded from the internet can trigger "Windows protected your PC". Click More info and then Run anyway, or before extracting, right-click the .7z, open Properties, tick Unblock and click Apply.

3. Run it

   Double-click run_nvidia_gpu.bat. A console window opens, ComfyUI starts, and your browser opens at 127.0.0.1:8188 about fifteen seconds later.

   The other .bat files are for particular situations:

   • run_nvidia_gpu_fast_fp16_accumulation.bat adds --fast fp16_accumulation, which trades a tiny amount of precision for speed on supported cards.
   • run_cpu.bat runs without a GPU. Fine for testing a small workflow, painful for anything else.
   • advanced\run_nvidia_gpu_disable_api_nodes.bat starts ComfyUI with the paid cloud API nodes switched off.

   

Route C: manual install from GitHub

The developer route: clone the repository, make a virtual environment, install PyTorch yourself. It's also the route on Linux, and the one to use if you want several versions side by side.

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI

python -m venv venv
venv\Scripts\activate

PyTorch first: this is the big one
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

pip install -r requirements.txt

python main.py

On Linux or a Mac, the only change is how you make the venv and switch it on:

python3 -m venv venv
source venv/bin/activate

The PyTorch line depends on your graphics card. AMD cards on Linux use pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2 instead. On an Apple Silicon Mac, the README says to install the latest PyTorch nightly first, following Apple's guide to accelerated PyTorch on Mac, then carry on from pip install -r requirements.txt.

Two things go wrong here, and both are easy to avoid:

• Installing the requirements before PyTorch. Install torch first, with the CUDA index URL. Otherwise pip fetches a CPU-only build and everything runs at walking pace.
• Forgetting to activate the venv. If you haven't run venv\Scripts\activate, you're installing into your system Python and you'll lose an evening working out why nothing matches.

There's also comfy-cli, which wraps all of the above in two commands:

pip install comfy-cli
comfy install

Route D: Comfy Cloud

The official hosted version runs the same ComfyUI on rented GPUs, with popular models already available. It costs money and your work lives on their servers, but if you're on a laptop with built-in graphics it's the difference between doing this and not doing it. The Desktop app can connect to it as well, from that first "How do you want to run Comfy?" screen.

Installing ComfyUI Manager, and the flag nobody mentions

ComfyUI Manager installs custom node packs, tells you what a downloaded workflow is missing, and keeps things updated. You'll want it by your second day.

It now comes with ComfyUI as a Python package, but outside the Desktop app it's not switched on by default. The Desktop app launches ComfyUI with --enable-manager for you. Everywhere else, it takes two steps.

On the portable build

Open the portable folder, type cmd in Explorer's address bar to open a command prompt there, and install the Manager's requirements into the bundled Python:

python_embeded\python.exe -s -m pip install -r ComfyUI\manager_requirements.txt

Then add the flag to the launcher so it sticks. Right-click run_nvidia_gpu.bat, choose Edit, and put --enable-manager at the end of the first line:

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-manager
echo If you see this and ComfyUI did not start try updating your Nvidia Drivers to the latest. If you get a c10.dll error you need to install vc redist that you can find: https://aka.ms/vc14/vc_redist.x64.exe
pause

Save it, close ComfyUI, and run the .bat again. The Manager is now in the app.

On a manual install

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

The flag is needed every time you start ComfyUI, so put it in whatever script or shortcut you launch it with. There are two related flags: --enable-manager-legacy-ui for the old interface, and --disable-manager-ui, which keeps the Manager's background jobs running but hides its interface.

What you get

Your first image

1. Open a template, not a blank canvas

   Use the Templates panel (or Workflow → Browse Templates). Start with something small, such as a Z-Image-Turbo or SD 1.5 text-to-image template, rather than a 50 GB video model.

   

2. Let it fetch the models

   If you don't have the files a template needs, it opens with a Missing Models panel that offers to download them, and shows how big each one is. Check the total. It's often 15 to 20 GB.

   

3. Know where files go if you download them yourself

   Downloading from Hugging Face or Civitai yourself is often quicker. Which folder a file goes in depends on what kind of file it is, and most official templates include a note telling you exactly where:

   

   The folders you'll use mostCopy

   ComfyUI/models/
   ├── checkpoints/        all-in-one models (SD 1.5, SDXL)
   ├── diffusion_models/   split models (FLUX, WAN, Qwen)
   ├── text_encoders/      clip_l, t5xxl, umt5, qwen
   ├── vae/                ae.safetensors and friends
   ├── loras/
   └── controlnet/

   Put the file in, then press R in ComfyUI or refresh the browser tab. It appears in the node's list without a restart.

4. Press Run

   The first run of a session loads the model into VRAM and is slow. The second run is the real speed. If a node turns red, the next-but-one section explains why.

One models folder for every install

The moment you have two installs, you don't want two copies of a 12 GB model. ComfyUI can read models from folders anywhere on your system, using a small config file.

In the ComfyUI folder of your portable or manual install there's a file called extra_model_paths.yaml.example. Rename it to extra_model_paths.yaml and point it at the folder you want to share. To reuse the Desktop app's models, point it at Desktop's shared folder:

comfyui:
    base_path: C:/Users/YOU/AppData/Local/Comfy-Desktop/ComfyUI-Shared/models/
    checkpoints: checkpoints/
    diffusion_models: diffusion_models/
    text_encoders: text_encoders/
    vae: vae/
    loras: loras/
    controlnet: controlnet/
    upscale_models: upscale_models/

Replace YOU with your Windows user name and restart ComfyUI. Both installs now see the same library. Two details: the Desktop app writes its own version of this file and marks it "do not edit manually", so make the change on the portable side; and a key can list several folders using the | syntax shown in the example file, which is how a folder like clip from an older setup can sit alongside text_encoders.

Keeping it updated

Desktop

It tells you. When an update is ready, a Desktop Update Ready button appears at the top of the window. To check yourself, open the menu, then Desktop Settings and the Updates panel, where you can also turn on automatic updates.

From inside ComfyUI

With the Manager switched on, Help → Update ComfyUI updates ComfyUI to the latest stable release and restarts it. The same menu shows what changed in recent releases.

Portable

There's an update folder next to the .bat files:

And once more: these reinstall requirements.txt, not manager_requirements.txt, so run the Manager's pip command again afterwards.

Manual

git pull
pip install -r requirements.txt
pip install -r manager_requirements.txt

When it doesn't work

A red node and "2 errors found"

Either the file isn't downloaded, it's in the wrong folder, or it's been renamed. Open the node's dropdown. If your file isn't in the list, ComfyUI isn't looking where you put it.

"Port already in use", or it won't start a second time

ComfyUI is already running: a second console, a Desktop instance, or an earlier run that didn't close properly. Close the old one, or start this one on a different port with --port 8189.

Device says "cpu"

PyTorch can't see your GPU. Update your NVIDIA driver first, which fixes it most of the time. On a manual install, you almost certainly have a CPU-only PyTorch; reinstall it with the CUDA index URL from Route C.

c10.dll, or another DLL error on startup

Install the Microsoft Visual C++ Redistributable from https://aka.ms/vc14/vc_redist.x64.exe. The portable build's own .bat file points you at this, because it's the most common startup failure on a fresh Windows machine.

Out of memory

Try these in order: lower the resolution; set the batch size to 1; use an FP8 or GGUF version of the model instead of FP16; close other programs using the GPU, including games and heavy browser tabs; and as a last resort, start ComfyUI with --lowvram.

A custom node pack broke everything

Rename its folder inside ComfyUI/custom_nodes/ by adding .disabled to the end, then restart. If ComfyUI comes back, you've found the culprit. The Manager's Conflicting view catches the other common cause: two packs providing the same node.

What to do next

Once you have a first image, the most useful next step isn't a bigger model. It's understanding the workflow you just ran. The workflow guide goes through every node in a real text-to-image workflow, and the workflow file is free to download, with a note beside each node.

Questions people ask