A step-by-step guide to installing and running ComfyUI on an Apple silicon Mac, with the Comfy Desktop app or a manual install, plus how much memory each model needs, where models go and how to fix common Mac errors.
ComfyUI runs on Apple silicon Macs (M1 or later) through Apple's Metal (MPS) backend. Intel Macs aren't supported by any of the official install paths.
The easiest way in is Comfy Desktop, a free app for macOS 13 or later that sets up Python, PyTorch and ComfyUI-Manager for you.
A manual install takes about six terminal commands and gives you more control over Python and PyTorch versions.
Memory decides what you can run: 16GB handles SDXL comfortably, and newer models like Flux, Z-Image or Krea 2 fit at that size only as compressed GGUF files.
ComfyUI can't run fp8 model files on a Mac yet, so download the bf16, fp16 or GGUF versions instead.
ComfyUI runs on any Mac with an Apple silicon chip, from the M1 onward. Comfy Desktop needs macOS 13 Ventura or later, and Comfy recommends at least 4.85GB of free disk space per install, before you download any models. A manual install follows Apple's PyTorch requirements, which call for macOS 14 or later.
Intel Macs are effectively out. Every official ComfyUI install path is Apple silicon only, and PyTorch stopped publishing Intel Mac builds after version 2.2, while ComfyUI needs 2.7 or newer.
Apple silicon Macs share one pool of memory between the CPU and GPU, so your total memory is what limits model size. ComfyUI doesn't publish an official table, but model file sizes give a good guide.
Memory | What runs well |
|---|---|
8GB | SD 1.5 models |
16GB | SDXL, plus Flux, Z-Image or Krea 2 as GGUF Q4 or Q5 files with a GGUF text encoder |
32GB | Z-Image at full bf16 precision, the Wan 2.2 5B video model, and Flux or Krea 2 as GGUF files |
64GB | Flux and Krea 2 at full precision, with Qwen-Image at full precision a tight fit |
96GB or more | Qwen-Image at full precision and larger video models |
The text encoders add up quickly. The SDXL base checkpoint is about 6.9GB on its own, but Z-Image Turbo in bf16 is 12.3GB plus an 8GB text encoder, Flux.1 Dev in bf16 is 23.8GB plus a 9.8GB encoder, and Qwen-Image in bf16 is 40.9GB plus a 16.6GB encoder.
There are two good ways to install ComfyUI on a Mac.
Comfy Desktop: a free app that installs everything in a few clicks, keeps several installs side by side and turns on ComfyUI-Manager for you. It's the route ComfyUI's own README recommends for Macs, and it's the best choice for most people.
Manual install: you clone ComfyUI from GitHub and set up Python yourself. It takes longer, but you choose your Python and PyTorch versions and can update ComfyUI the moment a new version lands.
There's no portable build for Mac. ComfyUI's portable downloads are for Windows only.
Download Comfy Desktop from comfy.org/download, open the .dmg file and drag the app to your Applications folder. When you launch it, create a new install, or add an existing ComfyUI install if you have one.
Each install comes with its own copy of Python and PyTorch built for Apple silicon, so you don't need to install either, and you don't need Git. If macOS blocks the app the first time you open it, go to System Settings, then Privacy & Security, and click Open Anyway.
Comfy Desktop keeps your installs in ~/ComfyUI-Installs and puts models and your input and output images in ~/ComfyUI-Shared, so every install can use the same models. Moving the app to the Trash doesn't delete those folders.
Open Terminal and run these commands one at a time. The first installs Apple's command line tools, which include Git.
1. xcode-select --install
2. git clone https://github.com/Comfy-Org/ComfyUI.git
3. cd ComfyUI
4. python3 -m venv venv && source venv/bin/activate
5. pip3 install --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cpu
6. pip install -r requirements.txt
The fifth command installs a nightly build of PyTorch, which is what ComfyUI recommends on a Mac. Despite the "cpu" in the address, these Mac builds include Metal support. ComfyUI's README says Python 3.13 is very well supported, and suggests 3.12 if a custom node gives you trouble.
To add ComfyUI-Manager, which installs custom nodes and models from inside ComfyUI, also run pip install -r manager_requirements.txt.
Comfy Desktop opens ComfyUI in its own window. For a manual install, go to your ComfyUI folder, activate the environment with source venv/bin/activate, then run python main.py, or python main.py --enable-manager if you installed ComfyUI-Manager. Then open http://127.0.0.1:8188 in your browser.
The quickest start is the Templates browser in the sidebar. Pick a workflow, and if its models are missing, ComfyUI lists them. In Comfy Desktop, Download all saves each file to the right folder. In a manual install, you download the files and move them yourself.
Models go in subfolders of ComfyUI/models:
checkpoints: all-in-one model files, such as SDXL or SD 1.5 checkpoints
diffusion_models: standalone diffusion models, such as Flux, Z-Image or Krea 2
text_encoders: the text encoders those models need
vae: VAE files
loras: LoRA files
Download .safetensors files, not .ckpt files, which can run code when they're loaded. On a Mac, skip any file labeled fp8, since ComfyUI can't run that format on Apple's GPU yet.
GGUF files compress a model's weights so it fits in less memory, at a small cost in quality. Install the ComfyUI-GGUF custom node through ComfyUI-Manager, put .gguf files in ComfyUI/models/unet, and load them with the Unet Loader (GGUF) node. A Q4 or Q5 file is the usual starting point on a 16GB Mac, and you'll want a GGUF text encoder too, loaded with the CLIPLoader (GGUF) node, since full-size text encoders alone can take 8 to 10GB.
Most first-run problems on a Mac come from memory or from model formats Apple's GPU backend doesn't support.
Black images: ComfyUI already applies a fix for a black-image bug in recent versions of macOS. If you still get black images, add the --force-upcast-attention or --fp32-vae launch flag, and remove --fp16-vae if you set it.
"MPS backend out of memory": lower the resolution or batch size, close other apps, or switch to a smaller GGUF version of the model. The --reserve-vram and --cache-none flags can also help.
"Trying to convert Float8_e4m3fn to the MPS backend" or "Undefined type Float8_e4m3fn": you've loaded an fp8 file. Download the bf16, fp16 or GGUF version instead.
"Not currently implemented for the MPS device": some operations don't run on Apple's GPU yet. Starting ComfyUI with PYTORCH_ENABLE_MPS_FALLBACK=1 runs them on the CPU instead, and --cpu runs everything on the CPU as a last resort.
Git or clone errors: run xcode-select --install. Comfy Desktop doesn't need Git.
Something broke after installing a custom node: start ComfyUI with your custom nodes disabled to see whether one of them is the cause.
In Comfy Desktop, launch flags go in the Startup Args tab of the Manage panel. In a manual install, add them after python main.py.
ComfyUI ships updates weekly and a major stable version roughly every two weeks, and many fixes for newer models and Macs arrive that way. Comfy Desktop builds from the latest stable version, and in a manual install you update by running git pull in the ComfyUI folder, then pip install -r requirements.txt.
Yes, on any Apple silicon Mac from the M1 onward. It uses Apple's Metal (MPS) backend for GPU acceleration. Intel Macs aren't supported.
Comfy Desktop, a free app for macOS 13 or later. It installs Python, PyTorch and ComfyUI-Manager for you and opens ComfyUI in its own window.
16GB handles SDXL comfortably and runs newer models like Flux and Z-Image as GGUF files. 32GB runs Z-Image at full precision, Flux and Krea 2 at full precision need about 64GB, and Qwen-Image at full precision is most comfortable with 96GB or more.
Recent ComfyUI versions fix the common macOS black-image bug automatically. If it still happens, start ComfyUI with --force-upcast-attention or --fp32-vae.
Not yet. ComfyUI can't run fp8 files on Apple's GPU, so use the bf16, fp16 or GGUF version of the model.
