What you are actually running

The foundations made the case that ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. is the most important piece of infrastructure in this field. This part is the manual: how to set it up, read a graph, build the core workflows, and drive it by hand or by machine. It is more hands-on than the rest of the book by design. If you only ever touch hosted tools you can skip it, but the operators and artists who get the most out of these models almost all end up here, because ComfyUI is where the whole pipeline becomes visible and editable rather than hidden behind someone else's button.

ComfyUI describes itself as a modular diffusion GUI, API, and backend with a graph interface, and that three-part description matters. The thing you install is not a web page, it is a local server that runs models and exposes them both to a browser and to an APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface.. It is maintained by Comfy Org, the company built around comfyanonymous, the developer who created the tool shortly after Stable Diffusion appeared in 2022. In September 2025 the project raised seventeen million dollars on the pitch of becoming, in the founders' phrase, the operating system of creative AI, and it ships on a weekly release cadence, which is worth remembering when a tutorial from six months ago no longer matches your menus.

There are three ways to install it. The desktop app, an Electron package for Windows and macOS, is the easiest path and bundles the server, the Manager, and a Python environment that sets itself up on first launch. The portable Windows build is a self-contained download you extract and run with no installation. The manual install is a git clone plus a PyTorch install for your hardware, which is what most serious users on Linux run. It supports NVIDIA, AMD, Apple Silicon, and Intel GPUs, wants a recent Python (3.13 is the recommended sweet spot) and PyTorch 2.4 or newer, and it runs fully offline by default, downloading nothing unless you ask. One detail captures the whole philosophy: its memory management is aggressive enough to run large models on GPUs with as little as one gigabyte of VRAM.