ComfyUI

Comfy-Org/ComfyUIReviewed by AI

What it does

ComfyUI runs open AI models on your own machine through a canvas of connected boxes: each box is a step (load a model, write the prompt, save the picture) and you join them into a workflow you can save and reuse. It handles image, video, audio, 3D and text models, plus inpainting, upscaling and masks. It runs offline and downloads nothing unless you ask. You supply the models yourself, as multi-gigabyte files placed in the right folders, and you need a reasonably capable graphics card. Windows and macOS have a desktop installer; Linux is a manual install.

How easy it is to install

B · installer plus one simple step · Windows

Windows B
macOS B
Linux D

Windows and macOS have the desktop installer the README calls "the easiest and best way to use ComfyUI for new users", so the work left is downloading model files into the right folders and having a graphics card that can cope. The Windows portable 7z in this repository is 1.4 to 1.9 GB and the README says it "is not recommended for regular users". Linux is harder: there is no installer, you clone the repository, choose and install the right PyTorch for NVIDIA, AMD ROCm or Intel, run `pip install -r requirements.txt` and start it with `python main.py`.

Numbers on GitHub

Stars: 144,364on Oct 31, 2026
Downloads: 10,566,875162,977 this week
Growth: +1,077stars in 5 days

Stars over time

Facts

Runs on: Windows, macOS, Linux
You need to install: Graphics card
To build from source: Python, Graphics card
Disk space: 1.9 GB–5.7 GB · from the installer
Accounts or keys: Comfy API (only for the optional partner nodes that call closed models)
Cost: Free · The program is free and GPL-3.0. Optional paid extras exist: the "API nodes" that reach closed models (switched off with --offline) and the company's paid Comfy Cloud for people without a suitable computer. The models themselves are usually free downloads, each with its own licence.
License: GPL-3.0 (Copyleft)
Can replace: Midjourney, Runway, Firefly, Photoshop
Last update: Nov 1, 2026

Risk signals (6)

Attentioncustom nodes are other peoples code running on your computer
The README advertises "Extend ComfyUI with custom nodes" and ComfyUI-Manager, "an extension that allows you to easily install, update, and manage custom nodes", enabled with `python main.py --enable-manager`. A custom node is Python code from a stranger that runs with your own permissions, and the manager's own flags mention "background features like security checks".
AI review
Attentiona shared image can carry a whole workflow
README, Notes: "Dragging a generated png on the webpage or loading one will give you the full workflow including seeds that were used to create it." Handy for copying other people's setups, but a workflow from the internet may point at custom nodes and files you have not inspected.
AI review
Infomodels are a separate multi gigabyte download
Nothing creative happens until you add models: "Put your SD checkpoints (the huge ckpt/safetensors files) in: models/checkpoints", "Put your VAE in: models/vae", and larger models "have multiple files" with their own subfolders. The install alone is estimated at 1945 to 5836 MB before any model.
AI review
Infoit is a web server on your machine
You start it with `python main.py` and use it in a browser; the README explains how to add a self-signed certificate (`--tls-keyfile key.pem --tls-certfile cert.pem`) and warns that one is "not appropriate for shared/production use". It also exposes "a local API for integrating workflows into applications", so it is worth not opening that port to other machines.
AI review
Infomoves fast and breaks extensions
README, Release Process: "Commits outside of the stable release tags may be very unstable and break many custom nodes", with a weekly release cycle, 164 releases and 5391 open issues. The interface itself lives in a separate repository and is merged in every two weeks.
AI review
Infoeach model has its own licence
The program is GPL-3.0, but the README lists dozens of third-party models to download (Flux, Qwen Image, Wan, HunyuanVideo and others) and partner nodes for closed models such as "Nano Banana, Seedance, Hunyuan3D". What you may do with the pictures and videos you make depends on each model's own terms, not on this licence.
AI review

Evidence is shown in its original language.

What Radar could not check

Whether your graphics card is enough: the README claims the biggest models run on 4 GB of video memory plus 8 GB of RAM, which is the project's own measurement
Real disk use once you add models, which can be far beyond the 1.9 to 5.8 GB of the install itself
How fast generation is on your hardware, and whether the "fastest local inference engine in the world" claim holds for you
What any custom node or downloaded workflow does on your computer
Whether the macOS and Linux routes work as smoothly as Windows, since only Windows files are published in this repository
What the optional paid API nodes and Comfy Cloud cost, which this repository does not state