ComfyUI installation can be simple or highly controlled depending on how you plan to use it. The official project now offers Comfy Desktop for supported operating systems, a Windows portable package, manual installation and command-line tooling.
This guide focuses on a clean setup that is easy to update and troubleshoot later.
Which installation method should you choose?
- Comfy Desktop: easiest for most supported desktop users. It manages the ComfyUI installation, Python environment and dependencies.
- Windows portable: useful when you want an isolated package that can live in its own folder.
- Manual installation: best for advanced users, Linux servers, custom Python environments and reproducible deployment.
- Comfy CLI: useful for command-line installation and dependency management.
Current Comfy Desktop requirements
The official Comfy Desktop documentation currently lists support for Windows 10+, macOS 13+ on Apple Silicon, and Debian-based Linux. Each standalone installation requires several gigabytes of disk space, and the docs recommend more RAM than the bare minimum for comfortable use.
Remember that ComfyUI itself is only part of the storage requirement. Model checkpoints can add many more gigabytes.
Install with Comfy Desktop
For most beginners, this is the cleanest route:
- download Comfy Desktop from the official ComfyUI site;
- install the launcher for your operating system;
- create a ComfyUI instance;
- let the app manage the Python environment and dependencies;
- launch the instance and run a known-good template before installing custom nodes.
The advantage is isolation: each instance can keep its own ComfyUI version, settings and custom nodes.
Manual installation
The official manual-install flow is:
- create a virtual environment;
- clone the ComfyUI repository;
- install the GPU-specific PyTorch stack;
- install ComfyUI dependencies;
- start ComfyUI.
A basic repository setup looks like:
git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt
python main.pyThe correct PyTorch command depends on your GPU backend, so use the current official ComfyUI or PyTorch instructions rather than copying an old CUDA or ROCm command from a forum post.
NVIDIA
NVIDIA is the most broadly supported path for many ComfyUI workflows. Install a PyTorch build compatible with your current driver/CUDA environment, then install the ComfyUI requirements.
AMD
AMD support uses ROCm on supported systems. Hardware and operating-system support changes over time, so verify the current ROCm matrix before committing to a specific build.
Apple Silicon
On Apple Silicon, ComfyUI uses Apple-compatible PyTorch acceleration. The official manual documentation includes a dedicated Mac ARM path, while Comfy Desktop supports modern Apple Silicon Macs directly.
Where do models go?
ComfyUI uses dedicated model directories such as:
models/checkpointsmodels/lorasmodels/vaemodels/controlnetmodels/diffusion_models
Different model families may require additional folders for text encoders, CLIP Vision or other components.
Share models between multiple interfaces
If you already have a large model library, you do not need to duplicate every file. ComfyUI supports extra model paths through extra_model_paths.yaml. This can point multiple ComfyUI instances, or ComfyUI plus another interface, at one central model folder.
Install custom nodes carefully
Do not install dozens of custom-node packs before confirming the base installation works. First run a standard workflow. Then add one extension at a time, preferably through a known source or ComfyUI-Manager.
Custom nodes execute code inside your environment, so treat them as software dependencies.
How to update ComfyUI
For a manual Git installation, the official documentation recommends pulling the latest code and reinstalling requirements inside the correct virtual environment:
cd ComfyUI
git pull
pip install -r requirements.txtFor important production workflows, keep a backup or snapshot before major updates so you can return to a known-good environment.
First-run checklist
- launch ComfyUI without errors;
- open the browser interface;
- load a simple official/template workflow;
- confirm the required model files are detected;
- generate one output successfully;
- save the workflow;
- only then add custom nodes or complex video pipelines.
Common problems
Model not found
Check the folder, file type and extra-model-path configuration, then restart ComfyUI after changing paths.
CUDA or ROCm errors
These usually point to a mismatch between drivers, PyTorch build and hardware support. Verify the current backend installation rather than repeatedly reinstalling ComfyUI itself.
Out of memory
Lower resolution, batch size or frame count, use an appropriate precision/quantized model, or use supported offloading options. Video models are particularly memory intensive.
Missing custom nodes
A downloaded workflow may depend on third-party nodes. Identify the missing nodes, install only those you actually need and restart the application.
Next steps
Once the installation is stable, continue with the main ComfyUI guide and ComfyUI Video Workflows.
