ComfyUI is a free, node-based app for running image and video generation models, such as Stable Diffusion and Flux, on your own hardware. Once it is set up, your prompts and images stay on your PC and there is no per-image charge. The hard part used to be installing it. In 2026 there are three official ways to do it on Windows, and the right one depends on your graphics card.
Forge PC has not benchmarked ComfyUI on our own hardware for this article. The requirements and steps below come from the official ComfyUI documentation, read on 13 November 2026. ComfyUI changes quickly, so check the docs if a menu looks different.
Which install should you choose?
Version | Best for | GPU support on Windows (per Comfy docs) | Updates |
|---|
Comfy Desktop | Most people | NVIDIA or AMD recommended; runs without a dedicated GPU too | Stable releases, automatic |
ComfyUI Portable | NVIDIA users who want the newest features | NVIDIA GPUs or CPU-only | Latest commits |
Manual install | Advanced users, Intel Arc, unusual setups | NVIDIA, AMD, Intel and others | Whatever you pull from GitHub |
If you have never used ComfyUI, start with Comfy Desktop.
What you need
The Comfy Desktop requirements for Windows are:
OS: Windows 10 or later
Architecture: x64 or Arm64
GPU: a dedicated NVIDIA or AMD GPU is recommended for good performance, but not required
Disk space: at least 4.85GB per installation, before you download any models
That 4.85GB is just the app. Image models are separate downloads and can be large, so plan for plenty of free SSD space. Video memory is the real limit on which models you can run comfortably. Our guide to a PC for Stable Diffusion and Flux and how much VRAM you need cover the hardware side in detail.
The docs also recommend Google Chrome version 143 or later for the interface, because earlier versions have known visual glitches and performance problems with ComfyUI.
Option 1: Install Comfy Desktop (recommended)
Download the Windows installer from ComfyUI's official download page. It detects your platform automatically.
Run the downloaded .exe installer and follow the prompts.
Launch Comfy Desktop from the Start menu or desktop shortcut.
On first launch, the setup asks how you want to run Comfy: on Comfy Cloud (hosted) or installed locally on your machine. Choose local if you want to use your own GPU.
Follow the rest of the first-run setup, then download a model and load a starter workflow.
Comfy Desktop is a multi-installation manager, so you can run more than one ComfyUI instance from the same launcher. If you used the older Legacy Desktop app, Comfy Desktop can detect and migrate it, carrying over custom nodes, workflows, models and settings. It backs up your legacy state first, but it reuses your existing data folder, so make your own backup if you want an untouched copy.
About automatic updates
The docs say Comfy Desktop checks for updates shortly after it starts and every 10 minutes while it runs, and downloads what it finds. Neither behaviour can be switched off. The only setting controls whether a downloaded update installs without asking. Automatic install is on by default, so turn it off if you want to approve updates yourself. Desktop installs track the stable ComfyUI release by default.
Option 2: ComfyUI Portable (NVIDIA or CPU)
The portable version bundles its own embedded Python environment and uses the latest ComfyUI commits, so you get new features sooner. According to the docs, it is Windows-only and supports NVIDIA GPUs or CPU-only operation. Because it is self-contained, you can keep it in its own folder on a large SSD. If you have an AMD or Intel graphics card, use Desktop or the manual install instead.
Option 3: Manual install (advanced)
The manual route supports every GPU type ComfyUI does, including NVIDIA, AMD, Intel and Apple Silicon. The docs' guidance for Windows is:
Python: 3.13 is recommended and very well supported. 3.14 works but some custom nodes may have issues, and 3.12 is a fallback if custom-node dependencies give you trouble.
NVIDIA: install stable PyTorch with CUDA 13.0. The docs say a cu130 or newer PyTorch build is required on NVIDIA 20-series cards and newer.
AMD on Windows: ROCm via AMD's multi-architecture PyTorch packages, supporting RDNA 2, RDNA 3, RDNA 3.5 and RDNA 4 GPUs. No separate HIP SDK installation is needed.
Intel Arc: native PyTorch support through torch.xpu.
CPU only: use the --cpu launch option, which the docs note is slower.
PyTorch 2.7 and newer is supported, but the docs strongly recommend using a newer version.
NVIDIA, AMD or Intel?
The official docs now cover all three GPU makers on Windows. NVIDIA cards are supported by every install route, including Portable. AMD Radeon RX 6000-series (RDNA 2) cards and newer are supported through ROCm. Intel Arc cards need the manual install. For a deeper comparison of the two big GPU makers for local AI, see NVIDIA vs AMD for local AI. For buying advice, see our best GPU for local AI guide.
Safety tips
Custom nodes run code. ComfyUI's power comes from community extensions, but install them only from sources you trust.
Models are large downloads. Keep them on a fast SSD with free space to spare.
Keep an eye on temperatures. Image generation can keep a GPU at full load for long periods. Our overheating guide explains what to check.
Our verdict
For most Windows users, Comfy Desktop is the way to install ComfyUI in 2026. Run the official installer, choose a local installation on first launch, and decide whether you want updates to install automatically. NVIDIA owners who want the newest features can use the Portable build instead. AMD RDNA 2-to-4 and Intel Arc users get official support through the manual route. Budget at least 4.85GB for the app, plus plenty of SSD space and VRAM for models. We have not benchmarked ComfyUI ourselves; everything here comes from the official ComfyUI documentation.
