ComfyUI Setup Guide: Desktop, Portable or Manual Install

ComfyUI is the open source node graph engine for running image, video, audio and 3D models on hardware you own, and there are three official ways to install it: a desktop app, a Windows portable archive, or a manual clone. It is back near the top of GitHub's daily trending page with about 365 new stars in a day against 125,786 total, and the desktop route gets a first generation running in roughly ten minutes.

• The project now lives at Comfy-Org/ComfyUI; the old comfyanonymous/ComfyUI URL 301 redirects there, so old bookmarks and clone commands still work.
• The README's own recommendation for new users is the desktop app on Windows and macOS. The portable build is explicitly "not recommended for regular users".
• Manual install is the only path that covers every OS and every GPU vendor, including AMD, Intel Arc, Apple Silicon and Ascend NPUs, and it is the one where you pick your own PyTorch build.
• Latest stable is v0.31.0, published 8 August 2026, on a roughly weekly release cadence targeting Mondays.

What is ComfyUI and why is it trending?

ComfyUI started in January 2023 as a node graph front end for Stable Diffusion and has since turned into something closer to a general media runtime. The README's own feature list now spans image generation, image editing, video, audio, 3D, segmentation and depth estimation, plus text generation with multimodal input. Same canvas, same queue, wildly different model families.

RelatedOpenWork Setup: Open-Source Claude Cowork in 15 Minutes

Two things are keeping it on the trending list. The first is release velocity tied to model releases: v0.28.0 landed 15 July, then v0.29.0, v0.29.2, v0.30.0 and v0.31.0 all shipped inside the following three weeks. When an open weights video model drops, ComfyUI templates for it tend to appear the same day. The second is organisational: the repository now sits under the Comfy-Org account, with comfyanonymous/ComfyUI serving a 301 redirect, and the project has split into a core repo, a separate frontend repo, and a desktop build.

The practical pitch is that nothing leaves the machine. The README is unusually explicit about it: the core does not download anything unless you request it, and --disable-api-nodes turns off the optional paid API nodes so every built in feature stays offline.

How do you install ComfyUI on Windows or macOS?

Use the desktop app. The README calls it "the easiest and best way to use ComfyUI for new users", it is available for Windows and macOS, and it wraps a stable core release rather than the moving master branch. Grab it from comfy.org/download. The current desktop build is v1.0.38, published 9 August 2026.

Nothing to type, which is the point. If you want the terminal instead, the two paths below are both first party.

How do you install the Windows portable build?

The portable archive is a self contained folder: extract it and run. It ships Python 3.13 and PyTorch CUDA 13.0, and the README is blunt that it is "not recommended for regular users". Pick the archive that matches your GPU. From the v0.31.0 release, the Nvidia build is the popular one at roughly 2.0 GB, with AMD at about 1.7 GB and Intel at about 1.6 GB.

# Nvidia, 20 series and above
https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia.7z

# Nvidia 10 series and older, pytorch cuda 12.6 and python 3.12
https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_nvidia_cu126.7z

# AMD and Intel
https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_amd.7z
https://github.com/comfyanonymous/ComfyUI/releases/latest/download/ComfyUI_windows_portable_intel.7z

Extract with 7-Zip or with Explorer, then run it. Two documented snags: if the archive refuses to extract, right click the file, choose properties, and hit unblock. If it will not start at all, update your Nvidia drivers. Do not use the cu126 build on a 20 series or newer card; the README says that in capitals for a reason.

How do you install ComfyUI manually on Linux, macOS or Windows?

Manual install is the only route that covers every operating system and every GPU vendor, and it is what you want if you already maintain a Python environment. Python 3.13 is described as very well supported; 3.14 works but some custom nodes may have issues. On PyTorch, 2.7 is the minimum, a cu130 or newer build is required on Nvidia 20 series and above, and anything more than six months old should be updated.

Clone the repo first, then install the PyTorch build for your hardware before the requirements file:

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

# Nvidia, stable
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

# AMD on Linux, stable ROCm
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2

# Intel Arc, torch.xpu
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

# then, in every case
pip install -r requirements.txt
python main.py

AMD RDNA 3, 3.5 and 4 owners on Windows have an experimental route through AMD's own nightly index, one URL per architecture:

# RDNA 3, RX 7000 series
pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx110X-all/

# RDNA 3.5, Strix Halo / Ryzen AI Max+ 365
pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx1151/

# RDNA 4, RX 9000 series
pip install --pre torch torchvision torchaudio --index-url https://rocm.nightlies.amd.com/v2/gfx120X-all/

Apple Silicon is the same manual flow with one substitution: install the PyTorch nightly following Apple's own Metal guide, then the requirements file, then python main.py. Older AMD cards that ROCm does not officially support have a documented workaround, an environment variable that lies about the architecture:

# 6700, 6600 and maybe other RDNA2 or older
HSA_OVERRIDE_GFX_VERSION=10.3.0 python main.py

# AMD 7600 and maybe other RDNA3 cards
HSA_OVERRIDE_GFX_VERSION=11.0.0 python main.py

There is also a packaged CLI, comfy-cli, currently at 1.15.0 on PyPI and needing Python 3.10 or newer. It is two commands:

pip install comfy-cli
comfy install

How do you add custom nodes and share models?

ComfyUI-Manager is the extension that installs and updates custom nodes, and as of the current README it is opt in behind a flag rather than bundled on by default:

RelatedBlock Buzz Setup: Self-Host the AI Agent Workspace

pip install -r manager_requirements.txt
python main.py --enable-manager

If you already have another Stable Diffusion UI with tens of gigabytes of checkpoints, do not copy them. Rename extra_model_paths.yaml.example to extra_model_paths.yaml and point it at the existing folders. In the portable build that file sits in the ComfyUI directory. Otherwise the layout is conventional: checkpoints in models/checkpoints, VAEs in models/vae, textual inversions in models/embeddings.

What are the gotchas before you rely on it?

Custom nodes are arbitrary code. The extension ecosystem is most of why people choose ComfyUI, and every custom node you install runs with your permissions inside the same Python process. Manager makes installing them one click, which is exactly the problem. Treat a node pack like a dependency you actually audit, not like a browser extension.

Master is not stable. The README warns that commits outside stable release tags "may be very unstable and break many custom nodes". The portable build deliberately tracks the latest commits, so the fastest install path is also the one most likely to break a workflow you depend on. Stable tags land roughly every two weeks.

The PyTorch version is the whole ball game. Nearly every manual install failure traces back to it. A "Torch not compiled with CUDA enabled" error has a documented fix, pip uninstall torch followed by the correct index URL, and it is worth pinning that command somewhere before you need it.

Do not expose it casually. There is a documented TLS option using --tls-keyfile and --tls-certfile, but the README's own example generates a self signed certificate and says plainly that this is not appropriate for shared or production use. A local node canvas with filesystem access is not a thing to put on the open internet.

Disk fills faster than you expect. The portable archive alone is 1.6 to 2.0 GB before a single checkpoint. Modern video and image models routinely ship as multi file sets, and the README notes the larger ones have multiple files that each go in a different subfolder. Read the model card, not just the download button.

Our take

Install the desktop app first. Almost everyone reading a setup guide wants to see a generation happen, not debug a CUDA wheel, and the fastest route to a working graph is the one with no terminal in it. The manual path is genuinely better, but it is better in ways you cannot appreciate until you have hit its limits, and going there first mostly buys you an afternoon of PyTorch archaeology.

What makes ComfyUI worth the setup is not the node graph, which is polarising and always will be. It is that the same interface now runs image, video, audio and 3D models from a dozen different labs, on your GPU, with the weights on your disk, and it usually supports them the week they ship. That combination is rare and it is why a three year old project is still collecting hundreds of stars a day. Just remember that a node graph with a one click extension installer is also a code execution surface. Install the pack, read what it does, and keep the machine you run it on boring.

Setup steps verified against the project's own README, release assets and documentation on 10 August 2026. Star counts from GitHub trending and the GitHub API on the same date.