Introduction
If you're just getting into AI image generation, ComfyUI is the most powerful and flexible interface available — but setting it up on Windows can feel overwhelming. The node-based workflow system is incredibly powerful, but the installation process has changed significantly over the past year.
This guide walks you through every step of installing ComfyUI on Windows in 2026, from checking your hardware to running your first workflow. I'll also share the tips and tricks I wish I'd known before spending hours troubleshooting.
Before You Start: Hardware Requirements
ComfyUI is GPU-accelerated, so your graphics card matters. Here's what you need:
Minimum Requirements
- GPU: NVIDIA RTX 3060 (12GB) or better
- VRAM: 8GB minimum (12GB recommended)
- RAM: 16GB system memory
- Storage: 20GB free space (models take up most of this)
- OS: Windows 10 (21H2+) or Windows 11
Recommended Setup
- GPU: RTX 4070 Ti Super (16GB) or RTX 4090 (24GB)
- VRAM: 16GB+ for SDXL and FLUX models
- RAM: 32GB system memory
- Storage: NVMe SSD with 50GB+ free space
Pro tip: If you're buying a GPU specifically for ComfyUI, the RTX 4070 Ti Super at 16GB VRAM is the best price-to-performance ratio in 2026. Check current prices on Amazon or Newegg.
What About AMD or Mac?
ComfyUI runs best on NVIDIA GPUs due to CUDA support. AMD users can use DirectML, but expect significantly slower performance. Mac users should look into alternative setups for now.
Step 1: Install NVIDIA Drivers
Before installing ComfyUI, make sure your NVIDIA drivers are up to date.
- Download the latest Game Ready Driver from nvidia.com/drivers
- Run the installer and select "Express Install"
- Restart your computer when prompted
Important: Don't skip this step. Outdated drivers are the #1 cause of ComfyUI startup failures.
Step 2: Install Python
ComfyUI requires Python 3.10 or 3.11. Here's the easiest way to install it:
- Go to python.org/downloads and download Python 3.11
- CRITICAL: Check "Add Python to PATH" during installation
- Run the installer and choose "Install Now"
- Verify the installation by opening Command Prompt and typing:
python --version
Common mistake: Forgetting to check "Add Python to PATH" is the most common installation error. If python --version doesn't work, reinstall Python and make sure to check that box.
Step 3: Install ComfyUI
Now for the fun part — getting ComfyUI installed:
- Open Command Prompt and navigate to where you want to install ComfyUI:
cd C:\ComfyUI - Clone the ComfyUI repository:
git clone https://github.com/comfyanonymous/ComfyUI.git - Create a virtual environment:
python -m venv venvthenvenv\Scripts\activate - Install dependencies:
pip install -r requirements.txt
This last step can take 5-10 minutes depending on your internet speed. Grab a coffee.
Step 4: Download Models
ComfyUI doesn't come with models — you need to download them separately. Here's where to get started:
Essential Models to Download
- FLUX Klein v1 (best for portraits) — Download from HuggingFace, place in
ComfyUI/models/checkpoints/ - SDXL 1.0 (versatile all-rounder) — Download from HuggingFace, place in
ComfyUI/models/checkpoints/ - ControlNet models (for pose/edge control) — Download from the ControlNet repo, place in
ComfyUI/models/controlnet/
Model Directory Structure
ComfyUI/
“— models/
“— checkpoints/ <– Main models (.safetensors)
“— vae/ <– VAE files
“— controlnet/ <– ControlNet models
“— loras/ <– LoRA fine-tunes
“— clip/ <– CLIP models
“— input/ <– Place images here for img2img
“— output/ <– Generated images appear here
“— custom_nodes/ <– Install custom nodes here Storage note: A complete model setup (FLUX + SDXL + ControlNet + LoRAs) will use approximately 30-40GB of disk space. Make sure your drive has enough room.
Step 5: Run ComfyUI
With everything installed, launch ComfyUI:
python main.py The first launch may take a minute as ComfyUI downloads additional dependencies. Once loaded, you'll see a message like:
To see the GUI go to: http://127.0.0.1:8188 Open that URL in your browser and you'll see the ComfyUI interface.
Step 6: Load Your First Workflow
ComfyUI works with workflow files (JSON). Here's how to get started:
- Download a workflow file (like our FLUX Klein Portrait workflow)
- Drag and drop it into the ComfyUI browser window
- Click "Queue Prompt" to generate
Your first image will appear in the output section. Congratulations — you're now generating AI art!
Optimizing ComfyUI for Your Hardware
Once you have ComfyUI running, these settings will help you get the best performance:
For 8GB VRAM Cards (RTX 3060, 4060 Ti)
- Use FP8 models when available
- Keep resolution at 1024x1024 or lower
- Use the
--lowvramflag:python main.py --lowvram - Stick to SD 1.5 models for faster iteration
For 12-16GB VRAM Cards (RTX 3080, 4070 Ti Super)
- Use FP8 for SDXL, FP16 for FLUX
- Resolution up to 1536x1536 for SDXL
- Can run most ControlNet models
--medvramflag for large workflows
For 24GB+ VRAM Cards (RTX 3090, 4090)
- Full precision models for maximum quality
- Resolution up to 2048x2048
- Batch processing multiple images
- No memory flags needed
Installing Custom Nodes
Custom nodes extend ComfyUI's capabilities. Here's how to install them:
- Navigate to the
custom_nodesfolder - Clone a node repository:
git clone https://github.com/ltdrdata/ComfyUI-Manager.git - Restart ComfyUI
The most essential custom nodes for 2026:
- ComfyUI Manager: One-click node installation and management
- Impact Pack: Mass processing and face restoration
- ControlNet Preprocessors: Edge detection, pose estimation
- IP-Adapter: Reference image control
Where to find nodes: The ComfyUI community on GitHub has hundreds of community-created nodes. Always check the last update date — nodes that haven't been updated in 6+ months may not work with the latest ComfyUI.
Troubleshooting Common Issues
"CUDA out of memory" error
- Add
--lowvramor--medvramflag - Reduce resolution
- Close other GPU-intensive applications
- Consider using FP8 models
"Module not found" errors
- Make sure you activated the virtual environment
- Reinstall dependencies:
pip install -r requirements.txt - Check that Python path is correct
Black or blank images
- Check your workflow connections
- Verify the model loaded correctly
- Try a different seed value
- Make sure your positive prompt isn't empty
Hardware Upgrade Recommendations
If you're looking to upgrade your setup for better ComfyUI performance, here are my top picks for 2026:
Best Value GPU
RTX 4070 Ti Super (16GB) — The sweet spot for AI generation. Handles SDXL and FLUX comfortably.
Best Performance
RTX 4090 (24GB) — The king of local AI generation. Everything runs fast.
Budget Option
RTX 3060 (12GB) — The cheapest way to get into serious AI generation. 12GB VRAM beats the 8GB cards.
Cloud GPU Alternatives
If buying hardware isn't an option, cloud GPUs are surprisingly affordable:
- RunPod: RTX 4090 starting at ~$0.59/hour
- VastAI: RTX 4090 starting at ~$0.29/hour
- RunComfy: Managed ComfyUI hosting, no setup required
Read our full guide on the best cloud GPU providers for ComfyUI
Final Thoughts
Setting up ComfyUI on Windows in 2026 is much more straightforward than it used to be. The biggest hurdles — Python installation, driver issues, and model management — have all been smoothed out with better documentation and automated tools.
The key takeaways:
- Use an NVIDIA GPU with at least 8GB VRAM (12GB+ recommended)
- Always add Python to PATH during installation
- Start with FLUX Klein or SDXL models
- Install ComfyUI Manager for easy node management
- Join the ComfyUI Discord for community support
Once you've got ComfyUI running, the possibilities are endless. From portraits to product photography to full scene generation, ComfyUI gives you the most control over your AI image generation workflow.
Affiliate Disclosure: This post contains affiliate links. If you purchase a product through one of these links, I may earn a small commission at no extra cost to you. I only recommend products and services I've personally tested. Thank you for supporting ComfyLab!
Have questions about setting up ComfyUI on Windows? Drop a comment below or join our Discord community for real-time help.