Images
A beginner's guide to ComfyUI on Windows

Key takeaways
Windows users have three ways in. The desktop app is what ComfyUI's repository calls the easiest way to start, the portable build is a zip you unpack, and the manual install is for people who want to manage Python themselves.
A dedicated NVIDIA or AMD GPU is recommended but not required. Intel Arc cards are supported too, and a CPU-only run works but is slow.
Models live in subfolders of a models directory. A checkpoint goes in the checkpoints folder, and pressing R reloads the list.
Every basic text-to-image graph uses the same six nodes, starting with Load Checkpoint and ending with Save Image. Bigger workflows add nodes to that same chain.
Out-of-memory errors and missing nodes are the two problems beginners hit first, and both have documented fixes.
How do you install ComfyUI on Windows?
For most people, download the desktop app, run the installer, and pick a local install on first launch. ComfyUI's GitHub page calls the desktop app "The easiest way to get started," and it says of the portable build that "It is not recommended for regular users."
ComfyUI is a node-based tool for running image and video models on your own machine. You build a generation as a graph of connected boxes instead of filling in a single prompt form, which is more work at first and gives you far more control later.
Desktop app, portable build, or manual install
All three run the same ComfyUI. They differ in who handles Python and updates.
Option | Best for | GPU support | Updates |
|---|---|---|---|
Desktop app | Most beginners | NVIDIA or AMD recommended | Checked automatically, installed automatically by default |
Portable build | People who want a self-contained folder | Separate downloads for NVIDIA, AMD, and Intel | Run an update script in the update folder |
Manual install | People who already manage Python environments | NVIDIA, AMD, Intel, and more | You update it yourself |
The Windows desktop docs list Windows 10 or later on x64 or ARM64 and say "At least 4.85 GB recommended per installation." That figure covers the app and its environment. Models take far more room. Our storage guide has real file sizes.
The desktop installer is a standard .exe. Run it, launch ComfyUI from the Start Menu, and on first launch choose between Comfy Cloud and a local install on your machine. Pick the local one if you want your own GPU doing the work. The docs say desktop update checks can't be turned off, though automatic installation of updates can be.
The portable build comes as a file you extract with 7-Zip, the only requirement the page lists. Inside you'll find run_nvidia_gpu.bat, run_amd_gpu.bat, and run_intel_gpu.bat, and you double-click the one that matches your card. It also has its own update folder, with update_comfyui.bat for the latest commit and update_comfyui_stable.bat for the latest stable version.
The manual install is mostly for people who want a particular Python version or a shared environment. The system requirements page says Python 3.13 is "very well supported and recommended," and that 3.12 is a good fallback if custom node dependencies give you trouble.
Which GPU do you need?
ComfyUI works best on an NVIDIA card, and the docs point NVIDIA users to PyTorch with CUDA 13.0. The portable page lists a CUDA 13.0 build with Python 3.13 for "modern Nvidia GPUs (RTX series)" and a CUDA 12.6 build with Python 3.12 for "Nvidia 10 series and older GPUs."
AMD works on Windows. The system requirements list ROCm via AMD's multi-architecture PyTorch packages for RDNA 2, RDNA 3, RDNA 3.5, and RDNA 4 cards. The portable build ships an AMD package, and a manual install on Windows asks for Windows 11, a current AMD driver, and 64-bit Python 3.13. Intel Arc cards use native PyTorch with torch.xpu support, and the portable page lists an Intel download too.
With no supported GPU, the docs say to use the cpu parameter, marked "slower." The troubleshooting page is blunter and calls it "very slow but works with any hardware, only use as absolute last resort." The portable build includes a run_cpu.bat in its advanced folder for this.
How much video memory you need depends on the model. Our VRAM guide breaks it down by model, and a card with 8GB can still run SDXL. Mac users should read the Mac vs. PC comparison, since this guide covers Windows only.
Where models go
ComfyUI sorts models into subfolders by type. In a portable install the path is ComfyUI\models, and the docs' examples use checkpoints, loras, vae, and controlnet. A checkpoint is the main model file, and the first-generation tutorial tells portable users to move the file into ComfyUI_windows_portable\ComfyUI\models\checkpoints.
The desktop app keeps its installs and a shared library apart. Per the docs, installations sit under %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Installs, and shared models and outputs sit under %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared. To store models somewhere else on desktop, edit the model directories in the Storage tab of the capsule menu.
Portable and manual users can point ComfyUI at another drive by copying the example file to extra_model_paths.yaml in the ComfyUI folder, then restarting. After you drop a new file into a folder, press R to reload the model list.
Your first image
The docs' own first-generation tutorial uses Stable Diffusion 1.5, specifically the file v1-5-pruned-emaonly-fp16.safetensors from Hugging Face. If you'd rather start from a ready-made graph, the template route below handles the model downloads for you.
Download the file, put it in the checkpoints folder, and press R. Then build the graph on a blank canvas, working left to right.
Load Checkpoint: set ckpt_name to your model file.
CLIP Text Encode (Prompt): you need two, one for the positive prompt and one for the negative prompt. The tutorial's examples are "beautiful scenery nature glass bottle landscape, purple galaxy bottle" and "text, watermark."
Empty Latent Image: width and height set to 512, batch size 1.
KSampler: leave the defaults, which the tutorial lists as 20 steps and a cfg of 8.0.
VAE Decode: takes the sampler's latent output and the VAE output from Load Checkpoint.
Save Image: takes the decoded image.
Click Run or press Ctrl+Enter. The result shows up in the Save Image node. In the app, open Assets and choose the Generated tab. On portable and manual installs the file is written to ComfyUI/output with a name like ComfyUI_00001_.png, and the number climbs with each image.
What each node in the graph does
Larger workflows are this same chain with extra nodes spliced in, so it pays to learn this one first.
Load Checkpoint loads the model and hands out three things, the model itself, the CLIP text encoder, and the VAE. The two CLIP Text Encode nodes turn your words into something the model can use. Empty Latent Image sets the size of the canvas the model starts from. KSampler does the actual denoising, and its steps and cfg settings control how long it runs and how closely it follows the prompt. VAE Decode converts the sampler's compressed latent output into pixels you can see, and Save Image writes the file.
Extra nodes slot in between these. A Load LoRA node sits between the checkpoint and the rest of the graph, and our LoRA explainer covers its two strength settings. An Apply ControlNet node adds pose, depth, or edge guidance, which the ControlNet explainer walks through.
Templates and the Manager
ComfyUI ships with built-in workflow templates, listed in the interface docs as "Built-in workflow templates." You can reach them from the Templates icon in the sidebar or from the Workflow menu, and loading one prompts you to download any model files you're missing. Templates are the quickest route to a working graph for a newer model, since you skip the node wiring. The Model Library in the sidebar then shows the "Models detected under your models folder."
The other tool to learn early is ComfyUI-Manager, which the repository describes as "an extension that allows you to easily install, update, and manage custom nodes for ComfyUI." Its docs list missing node detection among its features, and it installs the nodes a downloaded workflow needs. The new Manager interface is "enabled by default for Desktop users." On a portable install you turn it on yourself by installing the dependencies from manager_requirements.txt and launching with the enable-manager flag.
Custom nodes are code written by third parties, and ComfyUI's install guide is direct about that. It says "Only install custom nodes from trusted authors and those commonly used by the community" and warns that "unverified plugins may pose security risks that could lead to system compromise."
Common errors and how to fix them
Out-of-memory errors usually come from asking for too much at once. The troubleshooting page lists these fixes.
Lower the resolution or the batch size.
Close other applications that use RAM or video memory.
Launch with the lowvram flag, which it describes as "Low VRAM mode (uses cpu for text encoder)."
Use the reserve-vram flag to hold back memory for Windows itself, in gigabytes.
Smaller model files help too. The VRAM guide covers fp8 and GGUF versions, which trade a little quality for a smaller footprint.
Missing nodes show up when you open a workflow that uses custom nodes you haven't installed. Open the Manager and use its missing node detection. If a node still shows as missing after installing and restarting, the custom node troubleshooting page lists that exact symptom, and its fixes are to check Manager for an update, look for alternatives in the ComfyUI Registry, or remove or disable the problem node. Checking the startup log for "import failed" errors is the install guide's advice after any manual custom node install. When you hit a problem you can't pin down, the troubleshooting page says to "Disable all custom nodes" and test again, since most reported issues come from them.
A missing model is a different message. A notice offers a View details link, and the Missing Models list in the side panel has a Download button where one is available.
Where to learn more
The ComfyUI repository has the install instructions and the keyboard shortcuts. The docs at docs.comfy.org include model-by-model tutorials with example workflows you can drag onto the canvas. Drag a workflow PNG onto the canvas and ComfyUI rebuilds the graph, so you can open someone else's result and see how they set it up.
ComfyUI on Windows FAQ
Is ComfyUI free?
The desktop app and portable build are downloads from ComfyUI's own docs and repository, and a local install runs on your hardware. The docs also describe Comfy Cloud, a hosted option you can choose on first launch instead of a local install.
Can I run ComfyUI without a GPU?
Yes, with the cpu parameter or the run_cpu.bat script in the portable build's advanced folder. The docs mark it as slower, and the troubleshooting page calls it a last resort.
Do I need to install Python for the desktop app or portable build?
No. The portable build comes with its own Python, and the desktop app sets up its own environment. Python versions only matter for a manual install, where the docs recommend 3.13.
Where do my generated images go?
On portable and manual installs they land in ComfyUI/output. In the desktop app you can open Assets and then the Generated tab.
Why does ComfyUI say a model is missing?
The workflow references a file that isn't in the matching models subfolder. Download it, put it in the folder named in the notice, and press R to refresh. Templates and the Missing Models panel can download many of them for you.
