Menu
Open source

ComfyUI

4 min readintermediateUpdated 28 Sept 2026
1 · In one line

ComfyUI is a free, open-source app where you build AI image, video and audio generators by wiring boxes called nodes together.

1 · What it is

ComfyUI is a free, open-source program for making pictures, video and sound with AI models. Instead of one prompt box, it shows a canvas of boxes called nodes. Each node carries out a task, and wires carry results from one node to the next. The whole diagram is called a workflow.

Each node already has code behind it. So you build by connecting boxes, not by writing programs. A basic text-to-image workflow does five things. It loads a model. It turns your prompt into numbers. It cleans up random noise step by step. It turns the result into pixels. Then it saves the picture. When you want a simpler screen, App Mode can hide the graph behind a few controls.

The models are the files that do the real work. You download them separately and put them in a models folder. ComfyUI runs on Windows, Linux and macOS, and its code is shared under GPL-3.0, an open-source licence. A PNG image it makes can carry its whole workflow inside. Drop that image onto the canvas and the exact recipe comes back. That includes the seeds that were used.

2 · Why it exists

ComfyUI opens up the stages behind a single prompt box.

Hidden stepsMaking an image involves several separate stages, such as loading a model, turning text into numbers and removing noise.
LinksIn ComfyUI each stage is a node, and links join one node to the next.
RerunsComfyUI reruns only the parts of the workflow that changed.
3 · How it works

Follow one text prompt through a basic ComfyUI workflow.

A basic ComfyUI text-to-image workflow Five nodes in a row. Load Checkpoint loads the model from the models folder. CLIP Text Encode turns the prompt into vectors. The highlighted KSampler node removes noise step by step; an Empty Latent Image node below it sets the canvas size. VAE Decode turns the latent into pixels. Save Image writes the picture to the output folder. ONE TEXT-TO-IMAGE WORKFLOW · EACH BOX IS A NODE Load Checkpoint CLIP Text Encode KSampler VAE Decode Save Image Empty Latent Image loads the model from models folder your prompt becomes number vectors removes noise step by step latent becomes pixels picture saved to output folder sets the canvas size latent
Each box is a node. The highlighted one, the KSampler, is where the picture is actually made.
  1. 1 · loadA Load Checkpoint node loads the image generation model from the models folder.
  2. 2 · encodeA CLIP Text Encode node turns your prompt into lists of numbers the model can use.
  3. 3 · sampleThe KSampler node starts from random noise. It removes the noise over several steps, guided by the prompt.
  4. 4 · decodeA VAE Decode node turns the compact result into a normal picture made of pixels.
  5. 5 · saveA Save Image node shows the picture and saves it to the output folder.

Change one node and run again: only the changed part and what depends on it runs again.

4 · Where it's used
WhoWhat they askWhat it works with
Illustrator“Can I keep the same scene but try several prompt variations?”A saved workflow with a queue of prompt changes
Video hobbyist“Can I make a short video clip on my own computer?”A video workflow running on their own computer
App developer“Can my app send a job to a workflow and get the image back?”The local API that runs a saved workflow
Student“How did someone make this image, and with which seed?”The workflow stored inside the shared picture file
5 · What it solves, and what it doesn't
solves
  • It shows each stage of generation as a node you can inspect and rewire.
  • It can run fully offline on your own computer once you have the model files.
  • It saves whole workflows as files, so others can load and repeat them.
  • Community custom nodes add features beyond the built-in set.
doesn't solve
  • Most installs come without models; you download those separately.
  • Not every model file works out of the box in the built-in nodes.
  • Custom nodes come from many community authors, not the team that maintains the core nodes.
6 · Go deeper
Diffusion modelsStable DiffusionLoRAControlNetText-to-image models

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. repoComfy-Org/ComfyUI: The most powerful and modular AI engine for content creation, Comfy Org · read 28 Sept 2026
  2. repoComfyUI LICENSE, Comfy Org · read 28 Sept 2026
  3. docsGetting Started with AI Image Generation, Comfy Org · read 28 Sept 2026
  4. docsComfyUI Text to Image Workflow, Comfy Org · read 28 Sept 2026
  5. docsNodes, Comfy Org · read 28 Sept 2026
  6. docsModels, Comfy Org · read 28 Sept 2026
  7. docsCustom Nodes, Comfy Org · read 28 Sept 2026
Suggest a correction