How to Install ComfyUI on Windows in 2026 (Step-by-Step Guide)

ComfyUI has become the standard tool for serious AI image generation. Unlike Automatic1111 (which is more beginner-oriented but limited) or cloud tools like Midjourney (which charge per image and give you no control), ComfyUI runs locally on your own GPU, costs nothing per image, and gives you node-based control over every step of the generation pipeline.

It has a reputation for being hard to install. That reputation is outdated. In 2026, ComfyUI has an official Desktop app that installs in under 5 minutes on any Windows machine. This guide covers both methods — Desktop for beginners and the manual Git method for people who want more control.

---

What You Need Before Starting

For ComfyUI Desktop (recommended for most users):

• Windows 10 or 11 (64-bit)
• An NVIDIA GPU with at least 6GB VRAM (for SDXL and FLUX models) or 4GB for older SD 1.5 models
• 20–50GB free disk space for models
• No Python, no Git, no command line required

For the manual Git method:

• Python 3.10, 3.11, or 3.12 (do not use 3.13 — incompatible with some dependencies)
• Git for Windows
• NVIDIA GPU with CUDA support (CUDA 12.x recommended for RTX cards)
• Same disk space requirements

---

Method 1: ComfyUI Desktop (Easiest — Start Here)

ComfyUI Desktop is the official standalone app released in late 2025. It bundles everything you need — Python, PyTorch, CUDA — into a single installer. This is the method most Windows users should use in 2026.

Step 1: Download the installer

Go to the ComfyUI GitHub releases page and download the latest .exe installer for Windows. Look for the file ending in -windows-x64.exe. It will be around 1–2GB.

Step 2: Run the installer

Double-click the downloaded file. Windows Defender may prompt you — click "More info" → "Run anyway" (the app is safe; it just lacks a code-signing certificate on some builds). The installer wizard handles everything. Accept the defaults unless you want to change the install directory.

Step 3: First launch — choose your PyTorch backend

On first launch, the Desktop app asks you to select your GPU backend:

• CUDA (NVIDIA) — choose this for any NVIDIA RTX or GTX card
• DirectML — only if you have an AMD GPU
• CPU — very slow; only for testing on machines without a dedicated GPU

Select CUDA and click Continue.

Step 4: Download your first model

ComfyUI Desktop includes a built-in model manager. On the Models tab, you can download SDXL, FLUX.1-schnell (free and fast), or FLUX.1-dev (requires a Hugging Face account and agreement to the license). FLUX.1-schnell is the best starting point — it generates high-quality images in 4–8 steps with no API key needed.

After download completes, go to Workflows → New → Load Default to open the basic text-to-image workflow.

Step 5: Generate your first image

In the default workflow, find the "CLIP Text Encode" node (the text box on the left). Type your prompt. Click "Queue Prompt" in the top right. Your first image should render in 15–60 seconds depending on your GPU.

That is the full installation process for Desktop. If it works at this step, you're done.

---

Method 2: Manual Git Installation (Advanced)

Use this method if you want more control over the Python environment, want to install custom nodes manually, or prefer to keep ComfyUI updated with git pull instead of waiting for Desktop releases.

Step 1: Install Python

Download Python 3.11 from python.org. During installation, check the box that says "Add Python to PATH" — if you miss this, fix it in System Environment Variables later.

Verify: open Command Prompt and run python --version. Should return Python 3.11.x.

Step 2: Install Git

Download Git for Windows from git-scm.com. Use the default options. After installation, verify with git --version in Command Prompt.

Step 3: Clone the ComfyUI repository

Open Command Prompt and navigate to where you want to install (e.g., your D: drive):

cd D:\
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI

Step 4: Install PyTorch with CUDA

Go to pytorch.org. Select: Stable → Windows → Pip → Python → CUDA 12.4. Copy the install command, which will look something like:

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124

This step downloads about 2–3GB. It may take 5–15 minutes on a slower connection.

Step 5: Install ComfyUI requirements

pip install -r requirements.txt

Step 6: Place your model files

Download a model from Hugging Face (FLUX.1-schnell is free at black-forest-labs/FLUX.1-schnell). Place .safetensors checkpoint files in ComfyUI/models/checkpoints/. VAE files go in ComfyUI/models/vae/.

Step 7: Launch ComfyUI

python main.py

ComfyUI will print a local URL (usually http://127.0.0.1:8188). Open that in your browser. The interface is identical to the Desktop app.

To update later: git pull in the ComfyUI folder, then pip install -r requirements.txt again.

---

Installing FLUX Models

FLUX.1 is the highest-quality model family available for local use in 2026. There are two versions:

FLUX.1-schnell — Fast, free to use commercially, no account needed. 4-step generation. Available at: huggingface.co/black-forest-labs/FLUX.1-schnell. Download the flux1-schnell.safetensors file (~23GB).

FLUX.1-dev — Slower, higher quality, non-commercial license. Requires a Hugging Face account and license agreement. Same file size range.

Both models require at least 10–12GB VRAM for full precision. For 6–8GB GPUs: use the quantized flux1-schnell-Q4_K_S.gguf variants available in the community, which require the ComfyUI-GGUF custom node.

For SDXL (Stable Diffusion XL), models run well on 6–8GB VRAM and are much faster than FLUX but lower quality. The juggernautXL or RealVisXL checkpoints are popular starting points.

---

No GPU? Use a Cloud GPU Instead

Not everyone has a desktop with a 12GB+ GPU at home. If you're on a laptop, have an AMD GPU, or simply don't want to manage local hardware, cloud GPU options let you run ComfyUI remotely.

Ampere offers on-demand GPU instances billed by the minute — no subscription, no minimum commitment. You get a full Linux environment with an NVIDIA A100 or H100, SSH access, and a public URL to ComfyUI running in your cloud session. Rates start at around $0.30–0.80/hour depending on the GPU tier.

The workflow: spin up an Ampere instance, SSH in, install ComfyUI using the manual method above, run your session, and terminate when done. You only pay for the time the GPU is actually running.

This is a good option for occasional use (testing a new model, running a large batch job) without the upfront cost of a dedicated GPU.

---

FAQ

Q: Does ComfyUI work on AMD GPUs? A: Yes, but it requires the DirectML backend (Windows) or ROCm (Linux). Performance and compatibility are lower than NVIDIA CUDA. Most community guides and custom nodes assume CUDA. If you have an AMD GPU, the DirectML option in ComfyUI Desktop is the easiest path — just select it on first launch.

Q: What is the minimum GPU for ComfyUI to be useful? A: An RTX 3060 (12GB VRAM) runs most SDXL models well and can handle quantized FLUX models. An RTX 4070 (12GB) runs FLUX.1-schnell comfortably. Below 6GB VRAM, you're limited to older SD 1.5 models at reduced resolution, which is still functional but far behind the current quality standard.

Q: ComfyUI Desktop vs. manual Git — which should I choose? A: Desktop for beginners and anyone who doesn't want to deal with Python environments. Manual Git for developers who want to run custom nodes, write scripts that interface with ComfyUI's API, or stay on the absolute latest updates before Desktop catches up.

Q: Can I run multiple models without re-downloading everything? A: Yes. In both the Desktop app and the manual install, models are stored in the ComfyUI/models/ folder. You can download as many as you want — just switch between them in the workflow's "Load Checkpoint" node. Models are never deleted automatically.

Q: Is ComfyUI safe to use? I see a lot of warnings about unknown publishers. A: The core ComfyUI application from the official GitHub (comfyanonymous/ComfyUI) is safe and open-source. The "unknown publisher" warning on the Desktop installer is a Windows code-signing limitation, not a security issue. Custom nodes from third parties should be reviewed before installing — stick to nodes with active GitHub repositories and community usage before running unknown code.

---

Next Steps After Installation

Once ComfyUI is running, the most useful things to learn next:

1. Custom nodes via ComfyUI Manager — Install ComfyUI-Manager to get a one-click marketplace for 500+ community nodes. Essential for ControlNet, upscaling, inpainting, and video generation.
2. ControlNet — Control image composition using pose, depth, or canny edge maps. Turns ComfyUI from "generate random images" into "generate exactly the image I have in my head."
3. Workflow sharing — The community at civitai.com and openart.ai shares premade ComfyUI workflows as JSON files. Download a workflow file, drag it into ComfyUI, and it loads instantly with all nodes pre-configured.

If you're running into terminal or Python setup issues, our terminal setup guide and VRAM optimization guide cover the prerequisites in detail.