Installation: get the Windows portable package running in 5 minutes

If you searched for "ComfyUI installation", you may have been scared off by tutorials saying "first install Python, then PyTorch, then configure CUDA". Good news: ComfyUI Windows installation has a route that avoids all of that—the official portable package: a 7z archive with a self-contained Python environment and the PyTorch matching your GPU, ready to use after extraction. Once the download finishes, 5 minutes is enough; how long the download itself takes depends on your network speed.

This lesson walks you through the Windows portable package route: pick the right package → extract → double-click to launch → see the node canvas in your browser. If you want the desktop version, or you're a Mac / Linux user, head straight to alternative installations.

What you'll be able to do after this lesson

• Download the official ComfyUI portable package matching your GPU (NVIDIA / AMD / Intel each have their own file)
• Extract the 7z archive and launch ComfyUI with the correct .bat script
• Open http://127.0.0.1:8188 in your browser and see the node canvas
• Explain "ComfyUI GPU requirements": which package to download for which card, and what to do without a discrete GPU
• Upgrade the portable version using the official scripts in the update folder

Where you are stuck right now

• You want to try AI image generation, so you search "ComfyUI Windows installation", and every tutorial says "install Python → configure pip sources → install PyTorch → match CUDA versions"—discouraging before you even start
• You installed an all-in-one bundle before, its Python environment conflicted with other software, and even uninstalling didn't clean it up
• You downloaded ComfyUI but don't know which file to double-click; a black window flashes by, or the service starts but you can't find the interface

When to use this approach

When you need to:

• Get ComfyUI running on Windows as fast as possible: no installing Python, no configuring CUDA, no touching system environment
• A "green" installation: the entire program along with its standalone Python environment lives in one folder
• Use the version officially pre-compiled for your GPU (NVIDIA / AMD / Intel) directly

And you don't want to:

• Fiddle with matching Python versions against PyTorch and CUDA
• Leave a pile of environment variables and dependencies on your system
• Learn terms like conda and venv before you can start generating images

💡 ComfyUI's developers also offer a desktop version

The official also provides the Comfy Desktop version, which the official README calls "simplest for new users", with a graphical installer and automatic self-updates. The portable package and the desktop version run the same ComfyUI, so pick either one; for how to install the desktop version, see alternative installations.

🎒 Preparation before you start

• Finish the previous lesson What is ComfyUI: know what nodes, connections, and workflows are, so the interface won't confuse you after installation
• A Windows computer: recent versions of Windows Explorer can extract 7z directly; on older systems, install the free 7-Zip first
• Figure out your GPU brand (NVIDIA / AMD / Intel): if unsure, check "Display adapters" in Device Manager
• Browser: Google Chrome 143 or higher is recommended—the official explicitly states that 142 and below have known display and performance issues
• Leave plenty of disk space: the portable package is sizable after extraction, and each model later can easily be several GB
• A cup of water ☕

⚠️ Tips for downloading from mainland China

The installation package is hosted on GitHub Releases, and models mostly come from Hugging Face—the official documentation notes that users in mainland China may need to configure a proxy for stable access to GitHub. A slow download doesn't mean you did anything wrong.

System requirements: ComfyUI GPU requirements quick reference

The official hardware baseline is quite low: with asynchronous weight streaming, 4GB of VRAM + 8GB of RAM can run even the largest open-source models. GPU support by vendor:

Hardware	Support status
NVIDIA	Cards from the last 10 years are supported (GTX 900 series and up); the official recommends 30 series and above for best performance—the more VRAM, the better
AMD	The Windows package is based on ROCm, supporting RDNA 2 / 3 / 3.5 / 4 architectures, with no separate HIP SDK needed
Intel	Arc series (native torch.xpu support)
Apple Silicon	Supported (M series)—follow the Mac installation route; see Alternative Installation
No discrete GPU	can run in CPU mode (--cpu), just slowly

As for operating systems, Windows / Linux / macOS are all on the official support list, but the portable package is only released for Windows. You don't need to install Python either: the portable package comes with its own standalone Python (currently 3.13) and PyTorch (the NVIDIA package is currently the CUDA 13.0 version).

Core Idea

What the ComfyUI portable package is The official release bundles ComfyUI, a standalone Python environment (python_embeded), and the PyTorch matching your graphics card into a single 7z file. "Installing" just means extracting it; "launching" just means double-clicking the .bat script inside—it runs ComfyUI\main.py with the bundled Python and starts a local service on your computer.

Where the interface is. ComfyUI uses a "local service + web interface" structure: after launching, access http://127.0.0.1:8188 in your browser (127.0.0.1 means "this machine"). That black command-line window is the service itself—the official docs explicitly warn: don't close it while in use; if you close it, ComfyUI stops running.

The most critical step: pick the package matching your graphics card. The official release offers different packages for different GPUs—this is the only place in ComfyUI portable installation where it's easy to choose wrong. Here are the three local installation methods side by side:

Installation method	Who it's for	In one sentence
Windows portable package (this lesson)	Windows users who want to get running fastest	Extract and use; comes with Python and PyTorch
Comfy Desktop edition	Those who want a graphical installer and automatic updates	Officially described as the simplest for new users
macOS / Linux manual installation	Mac (Apple Silicon) and Linux users	git clone + pip, set up the environment yourself

Installed ≠ able to generate images. The installation package contains no model files—the first model (SD1.5, used in the official tutorial, about 4GB) must be downloaded and placed into the models directory yourself; that's covered in the Where to Put Models lesson.

Follow Along

Step 1: Confirm your graphics card and pick the right ComfyUI download package

Why: The official release packages separately by GPU type; picking the wrong package may cause launch failure or very poor performance.

Check the table below and click to download (all files are on the latest release page of GitHub Releases):

Your graphics card	Package to download	File name
NVIDIA RTX 20 series and above	Standard package (CUDA 13.0, Python 3.13)	ComfyUI_windows_portable_nvidia.7z
NVIDIA GTX 10 series and older	cu126 package (CUDA 12.6, Python 3.12)	ComfyUI_windows_portable_nvidia_cu126.7z
Windows on ARM + NVIDIA	ARM64 package (Python 3.14)	ComfyUI_windows_portable_nvidia_arm64.7z
AMD	AMD package (ROCm)	ComfyUI_windows_portable_amd.7z
Intel	Intel package	ComfyUI_windows_portable_intel.7z
No discrete GPU	Any NVIDIA package, then launch with run_cpu.bat	—

⚠️ Don't download the cu126 package for new cards

From the official README: the cu126 package is only for GTX 10 series and older graphics cards; do not use it on newer cards of the RTX 20 series and above.

You should see: your browser starts downloading a .7z file whose name matches the table above.

• The downloaded file name matches my graphics card

Step 2: Extract the 7z package—ComfyUI installation is now complete

Why: The portable package has no installer; "extracting" is the entire installation process.

What to do: Extract the .7z to a location with plenty of space (for example D:\). Right-click and extract with 7-Zip; newer versions of Windows Explorer can also extract directly.

You should see: an extracted folder with the following structure—this structure will be used in every lesson from here on:

ComfyUI_windows_portable
├── ComfyUI                   ← main program (the models directory, models, is also inside)
├── python_embeded            ← bundled standalone Python environment
├── update                    ← batch scripts for upgrading
├── README_VERY_IMPORTANT.txt ← official instructions for the portable package
├── run_cpu.bat               ← CPU mode launch (for no discrete GPU; slow)
└── run_nvidia_gpu.bat        ← NVIDIA launch (the AMD package uses run_amd_gpu.bat, the Intel package uses run_intel_gpu.bat)

• I can find run_nvidia_gpu.bat (or the .bat that matches my graphics card)

Step 3: Double-click the .bat to launch ComfyUI

Why: .bat uses its bundled Python to start a local service — this step is the actual "power-on".

What to do: Double-click the launch script that matches your graphics card:

• NVIDIA: run_nvidia_gpu.bat
• AMD：run_amd_gpu.bat
• Intel：run_intel_gpu.bat
• No dedicated GPU: run_cpu.bat

All it does boils down to one line — the line in run_nvidia_gpu.bat that actually does the work is this one (the file also contains an official troubleshooting note and pause; see "Pitfall Reminders" below for the note's content):

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

What you should see: A black command-line window scrolling logs, ending with:

To see the GUI go to: http://127.0.0.1:8188

Usually the browser opens automatically and navigates to this address; if it doesn't open automatically, open your browser yourself and visit http://127.0.0.1:8188 — once the node canvas appears, the installation succeeded.

⚠️ Don't close the black window

Official warning: do not close that command-line window while using it, otherwise ComfyUI will stop running. When you want to finish, close it and the service stops; next time, just double-click .bat again to restart.

• My browser opens http://127.0.0.1:8188 and I can see the node canvas

Step 4: Get to know the next destination — the models directory

Why: The installation package does not include models, so Load Checkpoint is still empty in the interface right now, and clicking Generate directly will show a red error. Installing models is the next lesson — for now, just learn where the door is.

What to do: Open ComfyUI_windows_portable\ComfyUI\models\checkpoints in File Explorer — this is where your first model goes (the official tutorial uses the SD1.5 weights v1-5-pruned-emaonly-fp16.safetensors).

What you should see: A row of subfolders under models (checkpoints, vae, loras, etc.). For a detailed explanation of the directory rules and how to share models across multiple programs, see Where to Put Models; for how to generate your first image after placing a model, see First Generation.

• I know my first model should go into ComfyUI\models\checkpoints

Checkpoints ✅

• The downloaded .7z filename matches my graphics card (I didn't mistakenly download the cu126 package for a newer card)
• After extraction I can see run_nvidia_gpu.bat (or run_amd_gpu.bat / run_intel_gpu.bat) and the update folder
• After double-clicking .bat, the command line shows To see the GUI go to: http://127.0.0.1:8188
• My browser opens http://127.0.0.1:8188 and I can see the ComfyUI node canvas
• I know: the black window must not be closed casually; installing isn't the end — models need to be downloaded separately

Pitfall Reminders

❌ After double-clicking the .bat, the black window flashes and disappears, and ComfyUI doesn't start

run_nvidia_gpu.bat contains a note written directly by the officials: if you see the note and ComfyUI doesn't start, first update your NVIDIA driver to the latest version; if you get a c10.dll error, install the Microsoft VC++ runtime library — the official link is https://aka.ms/vc14/vc_redist.x64.exe. Correct order: update the driver → then double-click; get c10.dll error → install the runtime library. For more startup issues, see What If It Won't Install.

❌ Downloaded the cu126 package for an RTX 30/40/50 series card

The cu126 package (CUDA 12.6) is intended for GTX 10 series and older cards; the official README emphasizes in capital letters "do not use on 20 series or newer". Correct approach: delete it and re-download the standard package ComfyUI_windows_portable_nvidia.7z (CUDA 13.0).

❌ Extraction fails with errors, or files are incomplete after extraction

Windows adds a "locked" mark to files downloaded from the internet, which can cause extraction to fail. The official README's fix: right-click the .7z file → Properties → check "Unblock" → OK, then extract again. If it still doesn't work, switch to 7-Zip for extraction.

❌ The browser can open 127.0.0.1:8188, but the interface is blank or garbled

The official system requirements state: Chrome 142 and earlier versions have known issues that cause interface display anomalies and performance problems. Correct approach: upgrade Chrome to 143 or higher before opening.

Advanced Tips (Optional)

🔧 Upgrading the portable version: The update folder contains three official scripts — just double-click them:

Script	Purpose
update\update_comfyui.bat	Update to the latest development version
update\update_comfyui_stable.bat	Update to the latest stable version
update\update_comfyui_and_python_dependencies.bat	⚠️ Reinstalls Python dependencies along with everything else; officially marked for use only when dependencies have problems — back up the entire installation directory before running

🔧 Letting other devices on the LAN access it: Open run_nvidia_gpu.bat with Notepad and add the --listen parameter (official example):

.\python_embeded\python.exe -s ComfyUI\main.py --listen --windows-standalone-build
pause

After restarting, the log will show To see the GUI go to: http://0.0.0.0:8188. Press Win + R, type cmd to open Command Prompt, and run ipconfig to find your machine's IP; other devices can then access http://<你的本机IP>:8188 in a browser.

🔧 Installing extra Python packages for the portable installation (custom node dependencies often need to be installed this way):

cd ComfyUI_windows_portable
.\python_embeded\python.exe -m pip install <包名>

If pip is slow in China, you can add the Tsinghua mirror parameter (this approach is also mentioned in the official docs): -i https://pypi.tuna.tsinghua.edu.cn/simple.

🔧 Official bundled launch variants: The NVIDIA package also includes a run_nvidia_gpu_fast_fp16_accumulation.bat, which enables fp16 accumulation acceleration — the official description is "faster for fp16 models, with slightly reduced quality"; the AMD package includes run_amd_gpu_enable_dynamic_vram.bat (with the --enable-dynamic-vram parameter).

Lesson Summary

✅ You have completed the ComfyUI installation (Windows portable package route): pick the package by graphics card → extract → double-click .bat → open http://127.0.0.1:8188 in a browser.

✅ You've figured out ComfyUI's GPU requirements: newer NVIDIA cards use the standard package, GTX 10 series and older use the cu126 package; AMD / Intel each have their corresponding packages; you can even run in CPU mode without a dedicated GPU, with the official minimum as low as 4GB of VRAM + 8GB of RAM.

✅ You've remembered three things: the black window is the service itself, don't close it carelessly; the installer doesn't include models, put the first model in ComfyUI\models\checkpoints; use the official script in the update folder to upgrade.

File / Address	Purpose
run_nvidia_gpu.bat	NVIDIA launch script (the AMD / Intel packages have their own corresponding scripts)
run_cpu.bat	Launch in CPU mode (add the --cpu flag, slow)
update\update_comfyui.bat	Upgrade to the latest dev version
ComfyUI\models\checkpoints	Put the first model here
http://127.0.0.1:8188	ComfyUI Interface address

Next Lesson Preview

The Windows portable package isn't the only way: Alternative installs: Comfy Desktop, macOS and Linux manual installation.

You will learn:

• How to install Comfy Desktop (officially described as the simplest for new users)
• Installation routes for Mac (Apple Silicon) and Linux
• How to choose among the installation methods, and whether you can switch after installing

Windows users who are already running the portable package can also jump straight to Where to put models to get their first image sooner.