ComfyUI
Want to create images or short videos with AI without paying per picture? ComfyUI lets you make them on your own computer by connecting ready-made steps instead of writing code.
How you use it: Download the desktop app, open a ready-made template, type what you want to see and run it.
In plain English
The README calls ComfyUI a modular engine for content creation. You build a workflow by connecting nodes instead of writing a program. It says the same engine can generate images, video, 3D, audio, and text, and that a desktop app, a portable install, and a paid cloud are all offered.
Local running is a major point. The README says it is tuned for consumer hardware and that even very large open models can run with as little as 4 GB of video memory plus 8 GB of RAM, using their weight-streaming approach. It also says the core stays offline unless you ask it to download something. A flag, --offline, disables optional paid API nodes.
The feature list names many open models for images, edits, video, and audio, plus partner nodes for closed models. Workflows save as JSON. The license GitHub reports is GPL-3.0. The repository now lives at Comfy-Org/ComfyUI.
Practical ways to use it
Example 1
Generate a picture from a written prompt
The README points at template workflows for text-to-image models. You open a template instead of wiring the first graph yourself.
Example 2
Edit an image you already have
The features section lists image editing workflows, inpainting, outpainting, and masks.
Example 3
Make a short video from an open model
The README lists video generation workflows, including named open models, as maintained templates.
Example 4
Work without a network
The README says the core does not download anything unless you request it, and --offline forces built-in features to stay offline.
Example 5
Call a closed model when you accept a paid node
Partner nodes are described as access to closed models. They are optional, and a disable flag turns those API nodes off.
Who it is for
Some tech comfort. The desktop app is a real application, and the node graph is a new skill. You do not have to write Python to use a template.
Difficulty: Takes some setup. The desktop download is the easy start on Windows and macOS. Manual install covers Linux and specific GPUs, and learning the graph takes practice.
Getting started
- 1
Download the desktop app if you are on Windows or macOS
The README calls the desktop application the easiest start and links www.comfy.org/download.
- 2
Or follow the manual install
The README says manual install supports Windows, Linux, and macOS, and GPU types including NVIDIA, AMD, Intel, and Apple silicon.
- 3
Open a template workflow
Use the template library linked from the README rather than a blank graph. Run one text-to-image template and wait for the model download it requests.
- 4
Stay offline if that matters to you
Pass --offline when you do not want paid API nodes, or the partner-node flag to turn off only those nodes. Use the cloud only if the hardware is not there, and check the site for the price.
What you need first
- Accounts
- None for local open models. Partner and API nodes, and Comfy Cloud, use accounts. The README treats those as optional.
- Hardware
- The README claims large open models can run on as little as 4 GB of video memory plus 8 GB of RAM. A desktop app exists for Windows and macOS. Linux uses the manual install.
- Cost
- Local open models are free to run aside from electricity. Partner nodes and Comfy Cloud are the paid paths the README names. Confirm prices on the site.
Cautions
- The node graph is powerful and easy to get lost in. Start from a template.
- GPL-3.0 applies to the project GitHub lists. If you ship software that includes ComfyUI, read that license with someone who knows it.
- Model downloads are large. The offline flags do not remove a download you already asked for. Watch disk space.
Questions people ask
What does ComfyUI do?
It is a node-based app for creating images, video, audio, 3D, and text with AI models, mainly on your own computer.
Who is ComfyUI for?
People willing to learn a visual graph. The desktop app lowers the install barrier. You do not have to be a programmer to run a template.
What do I need first?
The desktop app on Windows or macOS, or the manual install on Linux, plus enough disk space for a model. The README talks about 4 GB of video memory and 8 GB of RAM as a low end.
What should I watch out for?
Start with a template, know that the license is GPL-3.0, and turn on --offline if you do not want paid API nodes calling out.