Introduction

If you have ever thought, “I want a whole studio on my own machine that can generate high quality AI images and videos,” ComfyUI is that studio. Nodes, models, custom graphs, late-night experiments — all local, all yours.

The best of all: installing it in 2026 is far kinder than it used to be. Comfy Org now ships a proper Desktop launcher for Windows and Apple Silicon Macs, a self-contained Windows portable build for people who like a folder they can move, and a clean from-source path that Linux (and every power user) still loves.

This guide walks all three operating systems, picks the right path for your hardware, and gets you to http://127.0.0.1:8188 without the usual fog.

The Magic: Three Ways In

Think of ComfyUI as the engine and Comfy Desktop as the warm foyer that welcomes you in. Desktop installs the engine, the Python environment, and the updates for you. Portable is a packed suitcase: unzip, double-click, go. Manual install is the atelier you build yourself — more work, more control, the only official first-class path on Linux.

A quick map:

  • Comfy Desktop — easiest path on Windows 10+ and macOS 13+ (Apple Silicon). Tracks the stable ComfyUI release. Manager is already on. About 4.85 GB per installation, 8 GB RAM minimum, 16 GB happier.
  • Windows Portable — embedded Python 3.13, latest commits, NVIDIA / AMD / Intel / CPU builds. Great when you want a movable folder or bleeding-edge git. Regular newcomers should still start with Desktop.
  • Manual git + venv — Windows, macOS, and Linux. Required on Linux for the engine itself. Python 3.13 is the sweet spot; 3.12 is the fallback when a custom node sulks; 3.14 works but some nodes still trip.

Hardware in plain language: a dedicated GPU is recommended, not mandatory. NVIDIA 20-series and newer want PyTorch with CUDA 13.0. Older NVIDIA 10-series need the special CUDA 12.6 portable. AMD on Linux uses ROCm 7.2. Apple Silicon uses PyTorch MPS (not MLX). CPU mode exists; it is patient, not fast.

Use a current Chrome (143+) when you open the UI. Older Chrome builds have been glitchy with the graph editor.

Pro Tip

Install onto an SSD and keep checkpoints on the same fast drive. ComfyUI is a node graph, but every generation still shuffles large tensors on and off disk. A slow drive makes a fast GPU feel shy. Also: create one shared models folder early and point every install at it with extra_model_paths.yaml. Future-you will blow a kiss across the room.

Windows Install

Path A — Comfy Desktop (start here)

  1. Download the official installer from comfy.org/download. The page detects Windows.
  2. Run the .exe (NSIS installer) and follow the prompts. Admin rights are not required for the standalone install.
  3. Launch Comfy Desktop from the Start Menu.
  4. On the Welcome screen, create your first installation. Desktop downloads ComfyUI, Python, and dependencies into their own instance.
  5. Coming from Legacy Desktop? Let the new app migrate custom nodes, workflows, models, and settings. The old install stays as a backup.

Open the instance when it is ready. The UI lives at http://127.0.0.1:8188.

Path B — Windows Portable

Grab the archive that matches your GPU from the latest GitHub release:

  • NVIDIA 20-series and newer: ComfyUI_windows_portable_nvidia.7z (Python 3.13, PyTorch CUDA 13.0)
  • NVIDIA 10-series and older: ComfyUI_windows_portable_nvidia_cu126.7z (do not use this on 20-series+)
  • AMD: ComfyUI_windows_portable_amd.7z
  • Intel Arc: ComfyUI_windows_portable_intel.7z

Direct NVIDIA link:

https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z

Extract with 7-Zip or a recent Windows Explorer. If extraction fails, right-click the file → Properties → Unblock. Then double-click the matching launcher:

  • run_nvidia_gpu.bat
  • run_amd_gpu.bat
  • run_intel_gpu.bat
  • run_cpu.bat

Leave the console window open. Closing it stops the server.

To update later, use the scripts in the update folder (update_comfyui.bat for latest git, update_comfyui_stable.bat for stable).

Path C — Manual on Windows

Install Git and the Microsoft Visual C++ Redistributable. Python 3.13 is recommended. Then:

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
python -m venv venv
venv\Scripts\activate

NVIDIA (20-series and above):

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt
python main.py

If torch claims it was not compiled with CUDA, uninstall it and run the same install line again.

macOS Install (Apple Silicon)

Desktop is the golden path on M1 and newer, macOS 13 Ventura or later. Intel Macs should use the manual route and expect CPU-only performance.

Path A — Comfy Desktop

  1. Download the macOS build from comfy.org/download.
  2. Open the .dmg and drag Comfy Desktop into Applications.
  3. Launch from Launchpad or Spotlight.
  4. First launch may trigger a Gatekeeper warning. Go to System Settings → Privacy & Security and click Open Anyway.
  5. Create your first installation on the Welcome screen. Migration from Legacy Desktop works the same as on Windows.

The foyer is built. Walk in.

Path B — Manual on Apple Silicon

Git plus a virtual environment, then PyTorch with MPS:

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python main.py

Official docs still point Apple Silicon at the latest PyTorch nightly for best Metal acceleration. Confirm MPS is alive:

python -c "import torch; print(torch.backends.mps.is_available())"

You want True. ComfyUI uses the MPS backend here, not MLX.

Linux Install

Linux has no published Desktop installer. The engine itself is installed from source; that is the path most creators should take. Building Comfy Desktop from source is optional extra furniture.

Recommended — ComfyUI from source

On Ubuntu / Debian:

sudo apt update
sudo apt install -y git python3 python3-venv python3-pip
git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
python3 -m venv venv
source venv/bin/activate

NVIDIA:

pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt
python main.py

AMD (ROCm 7.2 stable):

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2
pip install -r requirements.txt
python main.py

Intel Arc (torch.xpu):

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu
pip install -r requirements.txt
python main.py

Headless or LAN access:

python main.py --listen 0.0.0.0 --port 8188

Shortcut — comfy-cli

If you already have Python and Git:

pip install comfy-cli
comfy install
comfy launch

Handy on Linux and anywhere you live in a terminal.

Optional — Comfy Desktop from source

Only if you specifically want the multi-instance launcher. Ubuntu 22.04+, x64, Node.js 22, pnpm 10+:

nvm install 22 && nvm use 22
corepack enable
git clone https://github.com/Comfy-Org/Comfy-Desktop.git
cd Comfy-Desktop
pnpm run init
./linux-dev.sh

If AppImages complain about FUSE: sudo apt install libfuse2. This is a development-style run, not a store installer.

First Launch, Models, and Manager

Drop checkpoints into models/checkpoints, VAEs into models/vae, LoRAs into models/loras. Bigger multi-file models have their own subfolders — follow each model card. Share one library across installs by copying extra_model_paths.yaml.example to extra_model_paths.yaml and pointing base_path at your shared drive.

ComfyUI-Manager:

  • Desktop: already included and enabled.
  • Portable: install manager requirements with the embedded Python, then launch with --enable-manager.
  • Manual: activate the venv, then:

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

Portable example:

.\python_embeded\python.exe -m pip install -r ComfyUI\manager_requirements.txt
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-manager

That Manager button is how you add custom nodes without turning your afternoon into a dependency novel.

A first-run checklist that never goes out of style:

  • GPU drivers current (NVIDIA especially — portable CUDA 13.0 is picky about stale drivers).
  • Browser pointed at http://127.0.0.1:8188.
  • At least one checkpoint in models/checkpoints.
  • Console or Desktop instance left running.
  • One shared models folder planned before you download 40 GB twice.

Conclusion

You now have a local generative studio on whichever machine you love most. Desktop if you want the door held open. Portable if you want the whole atelier in one folder. From-source if you like the smell of fresh git commits and a Linux terminal that already knows your name.

The graph is empty on purpose. Load a workflow, drop in a model, queue a prompt, and watch the first image bloom. That is the whole point of a local install: the lights stay on until you are done creating.

When you are ready for the next layer — latent upscalers, custom nodes, shared workflows — you already have the foundation. Come back to ArtRealmAI.com and we will keep the studio warm.

—Aurelia ♡