--- title: "Generate Images and Videos with ComfyUI — Image Gen Quick Start" canonical: "https://build.nvidia.com/spark/comfyui/image-gen-quick-start.md" --- > [!NOTE] > These instructions target **Linux**. This tab is a lightweight host Python install for Z-Image-Turbo text-to-image generation — a good first try on any supported hardware platform. For FLUX, Wan, HunyuanVideo, Cosmos, and playbook workflows, continue to the **Video Gen Workflow** tab. # Quick start (optional) If you prefer an automated setup, download and run the provided script to perform Steps 1–6 in one go (prerequisite check, virtual environment, PyTorch, ComfyUI, dependencies, and model download): ```bash curl -fsSL "https://raw.githubusercontent.com/NVIDIA/dgx-spark-playbooks/refs/heads/main/nvidia/playbook-comfyui/assets/setup.sh" | bash ``` When it finishes, launch the server from the **same directory** where you ran `setup.sh` (it expects `comfyui-env/` and `ComfyUI/` in the current directory): ```bash curl -fsSL "https://raw.githubusercontent.com/NVIDIA/dgx-spark-playbooks/refs/heads/main/nvidia/playbook-comfyui/assets/launch.sh" | bash ``` Then continue from [Step 8. Validate installation](#step-8-validate-installation). To learn what each step does, follow the manual instructions below instead. # Step 1. Verify system prerequisites Check that your hardware platform meets the requirements before proceeding with installation. ```bash python3 --version pip3 --version nvidia-smi ``` Expected output should show Python 3.8+, pip available, and GPU detection. # Step 2. Create Python virtual environment You will install ComfyUI on your host system, so you should create an isolated environment to avoid conflicts with system packages. ```bash python3 -m venv comfyui-env source comfyui-env/bin/activate ``` Verify the virtual environment is active by checking the command prompt shows `(comfyui-env)`. # Step 3. Install PyTorch with CUDA support Install PyTorch with CUDA 13.0 support. ```bash pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130 ``` This installation targets CUDA 13.0 compatibility with Blackwell architecture GPUs. # Step 4. Clone ComfyUI repository Download the ComfyUI source code from the official repository. ```bash git clone --branch v0.33.2 https://github.com/comfyanonymous/ComfyUI.git cd ComfyUI/ ``` # Step 5. Install ComfyUI dependencies Install the required Python packages for ComfyUI operation. ```bash pip install -r requirements.txt ``` This installs all necessary dependencies including web interface components and model handling libraries. # Step 6. Download the models that will be used in Step 9 Z-Image-Turbo ships as three separate files — a diffusion model, a text encoder, and a VAE — each in its own directory under `models/`. ```bash wget -P models/diffusion_models/ https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/diffusion_models/z_image_turbo_bf16.safetensors wget -P models/text_encoders/ https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/text_encoders/qwen_3_4b.safetensors wget -P models/vae/ https://huggingface.co/Comfy-Org/z_image_turbo/resolve/main/split_files/vae/ae.safetensors ``` The downloads total about 20 GB (12 GB diffusion model, 8 GB text encoder, 335 MB VAE) and may take several minutes depending on network speed. # Step 7. Launch ComfyUI server Start the ComfyUI web server with network access enabled. ```bash python main.py --listen 0.0.0.0 ``` The server will bind to all network interfaces on port 8188, making it accessible from other devices. # Step 8. Validate installation The server from Step 7 keeps running in the foreground, so run the following in a **second terminal**. Check that ComfyUI is running correctly and accessible via your web browser. ```bash curl -I http://localhost:8188 ``` Expected output should show HTTP 200 response indicating the web server is operational. Open a web browser and navigate to `http://:8188` where `` is your device's IP address. # Step 9. Run a template flow Test the installation with a basic image generation workflow: 1. Access the web interface at `http://:8188` > [!NOTE] > If the page fails to load, make sure the device you are browsing from is allowed to access your local network: > - **macOS:** Open **System Settings → Privacy & Security → Local Network** and enable access for your browser. macOS blocks local-network connections until an app is granted this permission. See [Control access to your local network on Mac](https://support.apple.com/guide/mac-help/control-access-to-your-local-network-on-mac-mchla4f49138/mac). > - **Windows:** Set your network profile to **Private** (not Public) so the device can reach others on the network. See [Make a Wi-Fi network public or private in Windows](https://support.microsoft.com/en-us/help/4043043/windows-10-make-network-public-private). 2. Load a starter workflow: 1. Click **Templates** on the left side of the menu (skip this if the template window pops up automatically) 2. Choose **Image** on the left side of the template window 3. Choose **Z-Image-Turbo: Text to Image**, the first template 4. Click the **Run** button at the top right 3. Monitor GPU usage with `nvidia-smi` in a separate terminal The image generation should complete within 30 seconds. > [!NOTE] > This is also how the Step 6 downloads were determined, and how to adapt these instructions to any other template. A template whose weights are missing reports each one as an error naming both the Hugging Face source and the `models/` subdirectory it belongs in (`diffusion_models/`, `text_encoders/`, `vae/`, and so on). Open the link in a browser, click **Copy download link**, and `wget` it into that directory. # Step 10. Optional - Cleanup and rollback If you need to remove the installation completely, follow these steps: > [!WARNING] > This will delete all installed packages and downloaded models. ```bash deactivate rm -rf comfyui-env/ rm -rf ComfyUI/ ``` To rollback during installation, press `Ctrl+C` to stop the server and remove the virtual environment.