Running ComfyUI locally: what it takes
Running ComfyUI on your own computer gives you full control, if your graphics card is big enough. This is what the setup involves, and what tends to go wrong.
Updated October 2026
What you need
- A graphics card with enough memory. 8 GB for SDXL-class pictures, 16 to 24 GB for current image models, and 24 GB or more for video. NVIDIA cards are the best supported. ComfyUI GPU requirements has the details.
- System memory. 16 GB at the least, 32 GB or more for big models and video.
- Disk space. A fast SSD with a few hundred gigabytes free. One video model with its text encoder can take 20 to 60 GB.
- Windows, Linux or an Apple silicon Mac. Macs run it more slowly, see ComfyUI on a Mac.
The ways to install it
- The desktop installer, for Windows with an NVIDIA card and for Apple silicon Macs. It sets up Python and the right graphics libraries for you. The easiest start.
- The portable Windows build, a folder with its own copy of Python. Extract it and run it, nothing else is installed.
- A manual install with Python, Git and the PyTorch version for your card (CUDA for NVIDIA, ROCm for AMD). Most control, most room for mistakes, and the usual route on Linux and with AMD cards.
Getting the models
ComfyUI comes without models. You download them, usually from Hugging Face, and put each file in the right folder. ComfyUI's built-in templates list the files each one needs and where they go.
| Folder | What goes there |
|---|---|
models/checkpoints | All-in-one model files (older models such as SDXL) |
models/diffusion_models | The main model file of newer image and video models |
models/text_encoders | The part that reads your prompt, often several gigabytes |
models/vae | The decoder that turns the result into pixels |
models/loras | LoRAs, small add-on files for a style or a character |
Many models come in several versions: full precision (bf16 or fp16), fp8 and GGUF. Pick the biggest one your card can hold.
Add-ons (custom nodes)
Workflows you find online often use add-ons that other people wrote, called custom nodes. ComfyUI Manager finds and installs them. Each add-on brings its own Python packages, and two add-ons that want different versions of the same package are the most common reason an install stops working.
What usually goes wrong
| What you see | What it means |
|---|---|
| CUDA out of memory | The model doesn't fit on your card. Use a smaller size, an fp8 or GGUF file, or the --lowvram setting. See the GPU requirements guide. |
| Missing node types (red boxes) | The workflow uses add-ons you don't have. Install them from ComfyUI Manager, then restart ComfyUI. |
| Torch not compiled with CUDA enabled | Python has the processor-only version of PyTorch. Install the CUDA version for your card, or use the desktop installer or portable build, which come with the right one. |
| A model doesn't show up in the list | It's in the wrong folder, or ComfyUI hasn't looked again. Move the file and refresh. |
| Everything broke after an update | An add-on needs a different version of a Python package than ComfyUI or another add-on. Update all add-ons, disable the newest one, or go back to the version that worked. |
| The first run is very slow | ComfyUI is reading tens of gigabytes from disk. Later runs are faster while the model stays loaded. A fast SSD helps. |
Your own card or a rented one
Running ComfyUI on your own card makes sense when the card is big enough and you enjoy building workflows: you have full control, nothing leaves your computer and each job costs only electricity. If your card is too small, or you'd rather not keep an install working, you can rent the card instead. Running ComfyUI without a good GPU compares the ways to do that, and ComfyUI without the setup compares running it yourself with CozyGPU.
Common questions
Is it hard to run ComfyUI locally?
With a supported NVIDIA card, the desktop installer gets ComfyUI itself running in minutes. Most of the work comes after: finding the right model files and add-ons for each workflow, putting them in the right folders, and keeping everything working after updates.
What do I need to run ComfyUI locally?
A graphics card with enough memory for the models you want (8 GB for SDXL-class pictures, 16 to 24 GB or more for current image and video models), 16 to 32 GB of system memory or more, and an SSD with a few hundred gigabytes free. NVIDIA cards are the best supported. Apple silicon Macs work, more slowly.
Can ComfyUI run offline?
Yes. Once ComfyUI, the add-ons and the model files are on your computer, it runs without an internet connection.
Where do ComfyUI model files go?
In the models folder inside ComfyUI: diffusion_models or checkpoints for the main model, text_encoders, vae and loras for the rest. A template or workflow tells you which file goes where.
Why does a ComfyUI workflow show missing nodes?
It was built with add-ons (custom nodes) you don't have installed. ComfyUI Manager can find and install them. Restart ComfyUI afterwards.
Is running ComfyUI locally more private?
Yes. Your prompts and results never leave your computer. A cloud GPU in your own account, which is how CozyGPU works, keeps them between your computer and that account, with no third service in between.
CozyGPU rents the graphics card for you, sets 12 open-source models up on it and gives you a simple app on your Mac or PC. It runs ComfyUI for you on a GPU in your own RunPod account, billed by the second, so nothing is charged while you're not using it. $20 a month, plus the GPU time.