How to Install ComfyUI on Windows in 2026 (Portable, Manual & Desktop App)

ComfyUI is the most flexible and powerful interface for running Stable Diffusion, FLUX, and other generative models locally on your Windows machine. If you’re learning how to install ComfyUI on Windows, you have three paths: the official Desktop app (fastest), a portable ZIP (no Python install needed, easy to move between drives), or a manual setup (most control). Unlike web-based tools, you get full control over your models, prompts, and workflows—and your images never leave your computer. This ComfyUI install guide covers all three so you can choose based on your comfort level and hardware.

At a Glance: Installation Methods Compared

The Desktop App is your move if you want to start generating images in minutes without touching the command line. Portable is the middle ground — no Python install, no terminal commands to launch it, and you can move the whole folder to another drive or PC without reinstalling. Manual install makes sense if you need multiple ComfyUI versions, plan to develop custom nodes, or want granular control over PyTorch and CUDA versions.

Option 1: ComfyUI Desktop App (Easiest)

The official ComfyUI Desktop application launched in January 2025 and handles most setup automatically. Most people should start here.

Download and Install

1. Visit github.com/Comfy-Org/desktop/releases
2. Download the latest Windows installer (.exe file)
3. Run the installer and follow the wizard—it detects your GPU and installs dependencies automatically
4. The app will prompt you to download a base model (SDXL or SD 1.5) on first launch
5. Launch the app and open the web interface in your browser

The desktop app needs roughly 15GB of disk space total, including bundled Python and PyTorch. It’s currently in beta but rock-solid for standard image generation workflows.

👉 Quick takeaway: ComfyUI Desktop is the fastest path to image generation—no command line required, automatic GPU detection, and you’re generating images within 15 minutes.

What You’re Trading Away

The desktop app handles 90% of use cases, but flexibility takes a hit. You can’t easily run multiple ComfyUI versions at the same time, and power users might find the UI-only update process limiting. Need full control or planning to build custom extensions? Skip to the manual install section.

💡 Tip: The Desktop app is perfect for testing whether ComfyUI fits your workflow before committing to a manual setup.

Option 2: ComfyUI Portable (No Python Install, No Terminal)

The portable build is a self-contained ComfyUI folder with its own bundled Python — you never touch a Python installer or a pip command.

1. Go to the ComfyUI GitHub releases page and pick the package matching your hardware: an Nvidia build (choose the one matching your CUDA version), an AMD (ROCm) build, an Intel build, or a CPU-only build if you don’t have a dedicated GPU.
2. Downloads come as .7z archives — extract with 7-Zip (or Windows’ built-in extractor on recent Windows versions).
3. Open the extracted folder. You’ll find a ComfyUI/ folder (the app itself), a python_embeded/ folder (the bundled Python — this is what skips the whole manual Python setup), an update/ folder with batch scripts for upgrading later, a README_VERY_IMPORTANT.txt, and launcher batch files like run_nvidia_gpu.bat, run_amd_gpu.bat, and run_cpu.bat.
4. Double-click the batch file matching your hardware. A command window opens and automatically launches your default browser to http://127.0.0.1:8188 once ComfyUI is ready.
5. Keep that command window open — closing it shuts down ComfyUI.

Model folders and everything else work exactly like the manual install described below — same models/checkpoints/, models/loras/, etc. structure inside the ComfyUI/ folder. Because it’s fully self-contained, you can move the whole extracted folder to another drive, or even another PC, and it keeps working without reinstalling anything.

💡 Tip: Add --listen to the batch file’s launch command if you want to reach ComfyUI from another device on your local network, and edit extra_model_paths.yaml if you want to point it at models stored elsewhere on your system instead of duplicating them.

Option 3: Manual Install (Full Control)

This takes longer but gives you complete control over every component. You’ll install Python, Git, ComfyUI, and PyTorch separately. Pick this route if you already have Python installed or want to customize your setup.

Step 1: Install Python 3.10 or 3.11

ComfyUI needs Python 3.10 or 3.11. Python 3.12 breaks several dependencies, so skip it.

1. Visit python.org/downloads and download Python 3.10 or 3.11 (3.11 is the safer bet)
2. Run the installer
3. Critical: Check the box labeled “Add Python to PATH” before clicking Install
4. Complete the installation
5. Open Command Prompt (Win+R, type cmd, press Enter) and verify:

   python --version

   You should see Python 3.10.x or Python 3.11.x. If Python isn’t recognized, reinstall and make sure that PATH checkbox was checked.

⚠️ Important: Missing the “Add Python to PATH” checkbox is the #1 reason people get stuck here. If you skip this step, ComfyUI won’t run.

Step 2: Install Git

ComfyUI lives on GitHub, so you need Git for cloning and updating.

1. Download from git-scm.com/download/win
2. Run the installer and accept all default options
3. Verify in Command Prompt:

   git --version

Step 3: Clone ComfyUI

1. Open Command Prompt and create a working directory:

   mkdir AI
   cd AI

2. Clone the repository:

   git clone https://github.com/comfyanonymous/ComfyUI

   Takes under a minute and creates a ComfyUI folder with all the source code.

Step 4: Create Virtual Environment and Install Dependencies

A virtual environment keeps ComfyUI’s Python packages separate from your system Python, preventing conflicts and dependency mess.

1. Navigate to ComfyUI:

   cd ComfyUI

2. Create the virtual environment:

   python -m venv venv

3. Activate it:

   venv\Scripts\activate

   You should see (venv) at the start of your command prompt line.

4. Install PyTorch with CUDA support—this is where most people stumble. First, check your CUDA version:

   nvidia-smi

   Look at the top-right corner of the output. You’ll see something like “CUDA Version: 12.1”. Most recent NVIDIA drivers (525+) support CUDA 12.1.

5. Install PyTorch matching your CUDA version. For CUDA 12.1 (the most common):

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

   For older drivers with CUDA 11.8:

   pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

   This download runs about 2GB and takes 5–10 minutes depending on your internet.

6. Install the rest of the dependencies:

   pip install -r requirements.txt

📌 Keep in mind: Mismatching your PyTorch CUDA version to your GPU driver is the #1 cause of “GPU not detected” errors. Get this right and everything else falls into place.

Step 5: Organize Your Model Folder Structure

ComfyUI expects models in specific subfolders. Create this structure inside your ComfyUI folder:

ComfyUI/
├── models/
│   ├── checkpoints/        (main models: SDXL, SD 1.5, FLUX)
│   ├── vae/                (standalone VAEs like ae.safetensors for FLUX)
│   ├── clip/               (text encoders: clip_l.safetensors, t5xxl for FLUX)
│   ├── loras/              (LoRA fine-tuning models)
│   ├── controlnet/         (ControlNet models)
│   ├── upscale_models/     (upscaling models)
│   └── unet/               (quantized GGUF models)

ComfyUI creates these folders automatically on first run, but you’ll need to populate them with actual model files.

Step 6: Download Your First Model

What you grab depends on your VRAM:

• 6GB+ VRAM: Download SDXL Base 1.0 (~7GB, produces higher quality). Available from Hugging Face or Civitai.
• 4–6GB VRAM: Download Stable Diffusion 1.5 (v1-5-pruned-emaonly.safetensors, ~2GB). It’s lighter and excellent for learning ComfyUI workflows.

Drop the downloaded .safetensors file into ComfyUI/models/checkpoints/.

Step 7: Launch ComfyUI

1. In Command Prompt (with venv activated):

   python main.py

   First launch takes 1–2 minutes as libraries load. You’ll see:

   To see the GUI go to: http://127.0.0.1:8188

2. Open that URL in your browser (Chrome, Firefox, Edge all work). The web interface loads.

Step 8: Create a Startup Batch File (Optional but Recommended)

Skip the command-line dance every time. Create a start.bat file in your ComfyUI folder:

@echo off
cd /d C:\AI\ComfyUI
call venv\Scripts\activate
python main.py --lowvram
pause

Use --lowvram for 4–6GB VRAM GPUs. If you have 8GB+ VRAM, drop the --lowvram flag. Double-click start.bat to launch ComfyUI from now on.

👉 Quick takeaway: A batch file eliminates command-line friction; save it and launch ComfyUI with a single double-click every session.

Step 9: Install ComfyUI Manager

ComfyUI Manager is the first extension you should grab. It lets you add and remove custom nodes from the UI without touching the terminal.

1. Close ComfyUI
2. Open Command Prompt and navigate to custom nodes:

   cd C:\AI\ComfyUI\custom_nodes
   git clone https://github.com/Comfy-Org/ComfyUI-Manager

3. Restart ComfyUI. A new “Manager” button appears in the menu bar.

Step 10: Generate Your First Image

1. In the web interface, confirm the CheckpointLoaderSimple node has your downloaded model selected
2. Click the CLIPTextEncode (Positive) node and enter a prompt:

   a beautiful landscape, golden hour, photorealistic

3. Click Queue Prompt (bottom-right button)
4. Your image generates in 20–60 seconds depending on GPU. It appears in the preview pane.

ComfyUI Desktop vs Manual Install: Side-by-Side Comparison

Troubleshooting Common Issues

”torch.cuda.is_available() returns False”

Your GPU isn’t being detected by PyTorch. Verify with:

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

If it prints False, your CUDA index URL was wrong. Update your NVIDIA drivers first, then reinstall PyTorch with the correct CUDA version.

CUDA Out of Memory

You’re running out of VRAM. Use:

python main.py --lowvram

Still not enough? Try:

python main.py --novram

This forces CPU processing (much slower, 5–10 minutes per image) but works on any hardware.

⚠️ Important: --novram mode is a last resort. If you’re hitting memory limits regularly, you may need a GPU with more VRAM or smaller models.

”Python is not recognized as an internal command”

Python isn’t in your system PATH. Reinstall Python and check the “Add Python to PATH” box during setup. Alternatively, manually add Python to Environment Variables via System Settings.

”Missing model error” or “Model not found”

The checkpoint filename in the CheckpointLoaderSimple node doesn’t match your actual file. Verify the .safetensors file exists in ComfyUI/models/checkpoints/ and matches exactly (including capitalization and extension).

Keeping ComfyUI Updated

ComfyUI gets updates regularly with performance tweaks and new model support. Update your manual installation:

cd ComfyUI
git pull
pip install -r requirements.txt

Run this every 2–4 weeks to stay current. The Desktop app checks for updates automatically.

FAQ

Q: What’s the minimum GPU I need to run ComfyUI? A: ComfyUI can run on NVIDIA GPUs with 4GB VRAM like the GTX 1650, though 6–8GB is recommended for SDXL models. With an RTX 3060 (12GB) you’ll get a smooth experience with most models.

Q: Can I use ComfyUI without a dedicated GPU? A: Yes, ComfyUI supports CPU execution, but generation will be much slower (minutes instead of seconds). For practical use, a CUDA-compatible NVIDIA GPU is recommended.

Q: What’s the difference between ComfyUI Desktop and the manual install? A: ComfyUI Desktop is an all-in-one installer that sets up Python, dependencies and the interface automatically. The manual install gives you more control and is better if you already have Python installed or want to customize the setup.

Q: Does ComfyUI work on Mac or Linux? A: Yes, ComfyUI runs natively on Linux and on Mac with Apple Silicon (M1/M2/M3). This guide focuses on Windows, but the process on Linux is similar using the terminal.

Keep Reading

New to ComfyUI entirely? Our What Is ComfyUI primer explains the node-based interface before you start building workflows. Once you’re up and running, the ComfyUI Manager guide covers adding custom nodes, and if PyTorch isn’t detecting your GPU, see the Torch not compiled with CUDA fix.

🏆 Our recommendation

If you’re new to ComfyUI and want results fast → Start with the Desktop App. You’ll be generating images in 15 minutes with zero command-line friction.

If you already have Python installed, want multiple versions, or plan to develop custom nodes → Go with the Manual Install. It takes 30 minutes but gives you complete control and understanding of how each component works.

If you’re switching between computers, don’t want to touch Python at all, or want to move ComfyUI to another drive without reinstalling → Go with Portable. It’s the least fuss of all three: extract, run a batch file, done.

Whichever path you choose, you’ll have a fully functional local generative AI setup within the hour. Download your first model, generate a test image, and explore the thousands of community workflows and custom nodes available. Your ComfyUI installation is ready.