Install ComfyUI (2026): The Setup Guide That Actually Works

To install ComfyUI, download Comfy Desktop from comfy.org/download, run the installer, confirm your GPU type, choose an install folder on an SSD and pick a starter template. Desktop bundles its own Python and ComfyUI-Manager, so on a supported Windows, Apple Silicon Mac or Linux machine there is nothing else to set up.

Checked October 2026 against Comfy's system requirements, Desktop for Windows, portable and manual install guides, the ComfyUI README and the latest release (v0.39.0, October 5, 2026).

This guide is built around what goes wrong. Each install path below lists its requirements, the steps, and the errors that path actually produces, with the fix. ComfyUI is the engine behind most of the workflows in our AI image generation for influencers guide; if you are not sure ComfyUI is the right tool yet, read what ComfyUI is first.

ComfyUI requirements: check these first

Comfy publishes no fixed VRAM minimum. The README says ComfyUI can run even the biggest open models on as little as 4 GB of VRAM plus 8 GB of RAM by streaming weights; more VRAM still means faster runs. If your machine falls short, our ComfyUI cloud vs local comparison covers renting a GPU instead.

Which ComfyUI install should you use?

How to install ComfyUI Desktop (recommended)

Desktop keeps shared models in %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared on Windows and ~/ComfyUI-Shared on macOS and Linux; add folders under Settings, Storage. To update ComfyUI, open Manage on the installation, then the Update tab (Stable is the default channel). See Comfy's instance management docs.

Desktop errors and fixes

• GPU not detected: update the GPU driver and rerun the installer; Desktop checks for a recent NVIDIA driver.
• Port busy: another ComfyUI is running. Close it, or use the If Port is Busy setting to pick the next free port.
• A template shows missing models: let Desktop download them, or put each file in the matching folder with its exact original name.

ComfyUI install on Windows: the portable build

The portable build is a folder with its own embedded Python. Download the archive for your GPU from the latest release:

• ComfyUI_windows_portable_nvidia.7z: GeForce 20 series and newer, with Python 3.13 and CUDA 13.0.
• ComfyUI_windows_portable_nvidia_cu126.7z: GeForce 10 series and older only, with Python 3.12 and CUDA 12.6.
• ComfyUI_windows_portable_amd.7z or ComfyUI_windows_portable_intel.7z for AMD and Intel GPUs.

Keep update_comfyui_and_python_dependencies.bat for repairing broken dependencies only, because it reinstalls every Python package. To turn on ComfyUI-Manager in the portable build, run these from the portable folder (also covered in our ComfyUI-Manager guide):

.\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

Manual install on Windows, Linux or macOS

You need Git and Python 3.13. On Windows, install the Microsoft Visual C++ Redistributable first. Comfy's guide uses Conda; a standard venv works too:

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
python -m venv venv
# Windows
venv\Scripts\activate
# Linux or macOS
source venv/bin/activate

Install the PyTorch build for your GPU before anything else, using the README commands:

# NVIDIA (CUDA 13.0)
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
# AMD on Linux (ROCm 7.2)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2
# Intel Arc (XPU)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

AMD on Windows uses AMD's ROCm 10.0 multi-architecture packages for RDNA 2 to RDNA 4; copy the exact command from the README, because its version pins change. On an Apple Silicon Mac, install the latest PyTorch nightly as the manual install guide describes. Then install ComfyUI's dependencies and start it:

pip install -r requirements.txt
python main.py

Expected result: the console prints the device it found and a local address; open http://127.0.0.1:8188. Check the GPU is in use with python -c "import torch; print(torch.cuda.is_available())" on NVIDIA. Prefer a single command? pip install comfy-cli followed by comfy install does the same job.

Every common ComfyUI install error and its fix

Find your message in the first column. The fixes come from Comfy's troubleshooting overview, model issues and custom node issues pages, and the README.

On memory errors: on NVIDIA GPUs with PyTorch 2.8 or newer, current ComfyUI turns on dynamic VRAM and logs "DynamicVRAM support detected and enabled". In that mode --lowvram does nothing, so go from --reserve-vram 2 straight to --novram, and use --cpu only as a last resort. Only use --listen on a network you trust, because it exposes ComfyUI to other devices.

Notes for Mac and Linux installs

Apple Silicon Macs. Desktop is the simplest route. ComfyUI does not use Apple's MLX framework; it runs PyTorch with the MPS (Metal) backend, according to Comfy's system requirements FAQ. Intel Macs are not a supported Desktop target, so they need a manual install and will be slow. Unified memory is shared between the system and the GPU, so close heavy apps before large generations.

Linux. Desktop ships as an AppImage or a .deb. On servers and headless machines, a manual install inside a venv is the usual choice: install the PyTorch build for your GPU, then requirements.txt, then start python main.py. To reach it from another machine, add --listen only on a private network, or tunnel the port over SSH instead of exposing it.

Reusing models from AUTOMATIC1111 or Forge

You do not need to copy gigabytes of models. In a portable or manual install, rename extra_model_paths.yaml.example in the ComfyUI folder to extra_model_paths.yaml, remove the # signs from its a1111 block, set base_path to your stable-diffusion-webui folder and restart. Desktop users add the same folders under Settings, Storage. Restart ComfyUI after adding any model file, because it scans folders at startup.

Updating, repairing and uninstalling

• Desktop: the app updates itself; update ComfyUI from Manage, then Update. Snapshots let you roll back if an update breaks a custom node.
• Portable: run the stable update script and keep the dependency script for repairs. To uninstall, delete the folder after copying out models and output.
• Manual: pull the latest code in the ComfyUI folder, then run pip install -r requirements.txt inside the same environment. To uninstall, delete the folder and its venv.

Before any update, note which custom nodes you rely on. Most post-update breakages are a custom node that has not caught up, and the --disable-all-custom-nodes test from the error table tells you within a minute whether that is the cause.

Generate your first image (the install test)

If that works, the install is done. Put files in the right folders from here: checkpoints for all-in-one models such as SD 1.5 and SDXL, diffusion_models for Flux, Qwen-Image and Wan, text_encoders, vae and loras. Download .safetensors where you can, since older .ckpt and .pt files can run code when loaded. Our best ComfyUI models guide lists what to download next.

For an AI influencer, the next steps are a consistent face and then motion: our face LoRA training guide covers the first. The AI Influencers course includes the ComfyUI workflows we use for persona images and video, ready to load once your install passes the test above.

Installing ComfyUI: FAQ