How to Install ComfyUI on Windows for Local AI Image Generation

Getting ComfyUI running on Windows takes two routes, and for a first install the choice is easy. The portable standalone build bundles its own embedded Python and pre-installed dependencies, so you download one archive, extract it, and launch. The manual route clones the repository and builds a Python virtual environment, which suits anyone who wants tighter control. Either way you end up with a node-based workflow graph in your browser, driving local AI image generation entirely on your own GPU.

Quick Answer

For beginners, use the portable build: download the Windows NVIDIA package from the official ComfyUI releases, extract it with 7-Zip, drop a model into the models folder, and run the GPU batch file. ComfyUI opens at a local browser address. You need an NVIDIA GPU with at least 6GB of VRAM, 16GB of system RAM, and plenty of free storage for models.

What You Need First

ComfyUI leans on CUDA, so an NVIDIA card is the smooth path. Six gigabytes of VRAM is the floor; 8GB is comfortable for SD 1.5, 12GB unlocks SDXL, and 16GB or more is what you want for heavier models like Flux. On the system RAM side, 16GB is the genuine minimum for a stable session, and 32GB is noticeably better when you keep other apps open alongside ComfyUI or load large models repeatedly. Storage is often underestimated: SDXL checkpoints run to several gigabytes each, Flux models are even larger, and a collection of LoRAs and VAEs can fill hundreds of gigabytes quickly. A dedicated NVMe drive of at least 1TB is the sensible starting point. If your current machine is short on any of these, a purpose-built AI PC takes the guesswork out of matching GPU, memory and storage.

Method 1: The Portable Build

1. Go to the official ComfyUI releases page and download the latest Windows NVIDIA portable package (the file is named along the lines of ComfyUI_windows_portable_nvidia).
2. Extract the archive with 7-Zip. Windows' built-in extractor works but is noticeably slower for a package this size.
3. Place at least one checkpoint model file into the models/checkpoints folder inside the extracted directory, since ComfyUI needs a model to generate anything.
4. Double-click the GPU batch file (run_nvidia_gpu) to launch. A console window opens and ComfyUI starts a local server.
5. Open the local browser address it prints, load the default workflow, type a prompt, and queue your first generation.

The portable build currently ships with a recent Python and a matching CUDA PyTorch, so you skip dependency wrangling entirely. If it fails to start, updating your NVIDIA driver is the first fix to try.

Method 2: Clone And Build A Virtual Environment

1. Install Git and a current Python release, then clone the ComfyUI repository to a folder of your choice.
2. Create a Python virtual environment inside that folder and activate it, which keeps ComfyUI's packages isolated from the rest of your system.
3. Install the CUDA build of PyTorch that matches your GPU, then install the project requirements.
4. Add your model files to models/checkpoints as before.
5. Run the main script to launch the server, then open the local address in your browser.

This route takes longer but makes updates and custom node management cleaner, which is why many regular users settle on it. Whichever build you pick, the heavy lifting falls on the GPU, and the cards SA creators reach for show up in the GPU best sellers.

First steps after launch

Once ComfyUI is running, load a text-to-image workflow, select your checkpoint from the model loader node, type a prompt, and queue the generation. The first run after launch is slower because the model loads into VRAM from scratch; subsequent runs on the same session are quicker because the weights stay resident. Start at a moderate resolution such as 512x512 or 768x768 to confirm everything works, then scale up once you know the setup is stable. ComfyUI Manager is worth installing early as it makes adding custom nodes and extensions straightforward from inside the interface.

Frequently Asked Questions

Do I need to install Python separately for the portable build?

No. The portable package includes its own embedded Python and dependencies, which is why it launches with a single batch file. The manual clone route is the one that needs a system Python.

How much VRAM is enough for ComfyUI?

Six gigabytes runs SD 1.5 workflows, 12GB handles SDXL comfortably, and 16GB or more is the target for Flux and other large models. More VRAM also means larger images and longer batches without running out of memory.

Why does my first launch fail to open?

The most common cause is an outdated GPU driver, so update it before anything else. A missing model file is the next suspect, since ComfyUI needs at least one checkpoint in the models folder to start generating.

Can I run ComfyUI on an AMD GPU?

It is possible on supported AMD cards, but NVIDIA's CUDA support is the gold standard and by far the smoother experience. For a first install, an NVIDIA GPU saves a lot of troubleshooting.

How much storage should I set aside for models?

Set aside at minimum 50GB for a starter library, though a working collection of checkpoints, LoRAs and VAE files grows past 200GB quickly. A dedicated NVMe drive makes loading noticeably faster than a hard drive.