What is ComfyUI
Making tomato scrambled eggs: prepare the ingredients, follow the steps one by one, and finally serve the dish. Each step does only one thing, and what the previous step produced is handed to the next.
Generating an image with AI follows the same kind of steps. First, line them up:
- 1?
- 2?
- 3?
- 4?
Think of cooking: first get the pot and ingredients, then cook, then plate. Tap in order.
ComfyUI is a free, open-source node-based AI creation engine. It breaks down "load the model, write the prompt, sample and generate, save" into individual nodes, strings them into a workflow with connections, and generates images, video, audio, and 3D on your own computer.
This section doesn't involve installing any software. After reading it, when you open it, the screen full of blocks and lines won't be intimidating anymore.
What you'll be able to do after learning
- Explain in one sentence what ComfyUI is
- Understand the 6 types of nodes in the official default workflow and the colors of the connections
- Answer three common questions: what can it do, is it free, and can my computer run it
Understanding it through a recipe
This analogy comes from the official basic concepts documentation:
| Cooking | ComfyUI |
|---|---|
| A recipe | A workflow |
| Each step in a recipe | A node |
| Ingredients | Prompts, images, models |
| The dish served at the table | The generated images, video, audio |
Behind each node is a program someone else has already written. You just drag, connect, and adjust parameters — no coding needed. The official term for this is "visual programming."
- I can explain what the workflow and the nodes each correspond to in a recipe
Connect a wire with your own hands
Here is the graph that ComfyUI usually loads automatically at startup: the official default text-to-image workflow. Most of the wires are already connected — four remain.
There's only one rule: only the same type of data can be connected; only ports of the same color fit together.
Load Checkpoint
CLIP Text Encode
CLIP Text Encode
Empty Latent Image
KSampler
VAE Decode
Save Image
Still need 4 Root line. The blinking port isn't connected yet: click an output port (on the node's right side), then click the same-colored input port (on the left).
- Model MODEL
- CLIP
- VAE
- Conditioning CONDITIONING
- Latent image LATENT
- Pixel art IMAGE
The four wires you connect are exactly the main arteries of the data:
- Load Checkpoint loads the large model and sends it out along three paths: MODEL, CLIP, and VAE.
- Two CLIP Text Encode nodes encode the positive and negative prompts into "conditions" the model can understand.
- Empty Latent Image prepares a blank latent-space canvas.
- KSampler gathers the four items — model, positive and negative conditions, blank canvas — and denoises step by step in latent space.
- VAE Decode decodes the latent-space data into an actual pixel image.
- Save Image saves it to disk.
| Node | What it does | Default parameters |
|---|---|---|
| Load Checkpoint | Loads the large model, outputs three paths: MODEL / CLIP / VAE | v1-5-pruned-emaonly-fp16.safetensors |
| CLIP Text Encode ×2 | Encodes the positive and negative prompts into conditions | Negative defaults to text, watermark |
| Empty Latent Image | A blank latent-space canvas | 512 × 512, batch 1 |
| KSampler | Sampler: denoises step by step in latent space to produce the image | Steps 20, CFG 8, euler / normal, denoise 1 |
| VAE Decode | Decodes latent-space data into a pixel image | None |
| Save Image | Saves to disk | Filename prefix ComfyUI |
The colors of the connections are officially defined; each color represents one type of data:
| Color | Data type | Who produces it | Who uses it |
|---|---|---|---|
| Light purple | Model (MODEL) | Load Checkpoint | KSampler |
| Yellow | CLIP | Load Checkpoint | CLIP Text Encode |
| Rose red | VAE | Load Checkpoint | VAE Decode |
| Orange | Conditioning (CONDITIONING) | CLIP Text Encode | KSampler |
| Pink | Latent image (LATENT) | Empty Latent Image, KSampler | KSampler, VAE Decode |
| Blue | pixels (IMAGE) | VAE Decode | Save Image |
No matter how complex a workflow gets later, this is how you read it: follow the colors to see where the data comes from and where it goes.
Quiz timeWhere should the pink line output by KSampler connect to?
- I connected the four lines in the diagram above and ran it once
- I know where the three outputs of Load Checkpoint each go
What it can do
According to the official feature list, node graphs cover five categories of creation:
| Capability | Description |
|---|---|
| Images | Text-to-image, image editing, inpainting, outpainting |
| Video | Text-to-video, image-to-video |
| Audio | Music and sound effect generation |
| 3D | 3D model generation |
| Text | Text generation (including multimodal input) |
It also comes with a set of processing tools: masking and compositing, model merging, upscaling, frame interpolation, segmentation, and depth estimation.
For supported models, the official README has a representative list; here is an excerpt:
| Task | Representative models in the official list | Corresponding tutorial on this site |
|---|---|---|
| Image generation | Stable Diffusion 1.5 / SDXL / SD3.5 / FLUX.1 / FLUX.2 / Qwen-Image / Z-Image | Z-Image-Turbo, Qwen-Image, FLUX.1, FLUX.2 |
| Image editing | FLUX.1 Kontext / FLUX.2 Klein / Qwen-Image-Edit | Qwen-Image-Edit, FLUX.1 Kontext |
| Video generation | Wan 2.1 / Wan 2.2 / LTX-Video 2 / LTX-Video 2.3 / HunyuanVideo 1.5 | Wan2.2 series, LTX-2/2.3 |
| Audio | ACE-Step 1.5 / Stable Audio 3 | None yet (see model overview) |
| 3D and vision | Hunyuan3D 2.1 / BiRefNet / SAM 3 | Product cutout with BiRefNet |
Auxiliary components like LoRA, ControlNet, VAE, text encoders, and upscaling models can all be loaded and combined. For the full map, see Model × Task Overview.
Is it free
The core is free and open source: the code is on GitHub under the GPL-3.0 license and runs on your own computer. The official docs state the core can run fully offline.
There are three things that cost money, all optional:
| Paid item | What it is |
|---|---|
| Comfy Cloud | The official paid cloud version, for those whose local hardware can't keep up (sign in with a Google account, with 400 free credits each month) |
| API / Partner nodes | Call closed-source cloud models within local workflows, billed by credits |
| Model files | Free to download, but a single model is often several GB, costing you disk space and bandwidth |
If you don't want to see any paid entry points, you can launch with the --offline argument.
Can my computer run it
The official threshold: with asynchronous weight streaming, 4GB of VRAM plus 8GB of RAM can run even the largest open-source models.
OS: Windows, Linux, macOS (Apple Silicon).
GPU: NVIDIA, AMD, Intel Arc, Apple Silicon, as well as NPUs like Huawei Ascend, Cambricon, and MetaX.
No dedicated GPU: add the
--cpuargument to run on the processor — it's just slow.I know the core is free, and which three things are optional paid items
I've checked it against my own computer
Compared with web-based AI image generation
| Web-based AI tools | ComfyUI | |
|---|---|---|
| Where it runs | The platform's cloud | Your computer (with an official cloud version as an option) |
| Cost | Per-use or subscription | Core is free and open source, unlimited local runs |
| Models | Whatever the platform offers | Pick open-source models yourself |
| Level of control | One prompt box | Every parameter and every step can be taken apart and modified |
| Workflow reuse | Hard | Workflows can be saved as JSON and are also embedded in generated images — drag one back onto the canvas to reproduce it |
| Offline | Unusable | The local core works fully offline |
The last row deserves a bit more explanation: ComfyUI-generated images carry the complete workflow information. When you see an image someone shared, drag it back onto the canvas and you've essentially got their recipe.
Checkpoints
- I can explain in one sentence what ComfyUI is: a free, open-source, locally run node-based AI creation engine
- I can list, in order, the 6 types of nodes in the default workflow
- I know that the color of a connection represents the data type, and only matching colors can be connected
- I know the core is free, while Comfy Cloud and API nodes are optional paid features
Common Misconceptions
Looking for a "Generate" button right after installing, only to get node errors. The installation package does not include any model files; you need to download models yourself. If you run it without placing any models, the Load Checkpoint dropdown list will be empty. First check where to put models.
Thinking you need to know how to code to use nodes. The official documentation states that using nodes requires no coding. When you need more features, use ComfyUI Manager to search for and install community custom nodes.
The interface is in English. You can switch it to Chinese in Settings → Comfy → Locale.
Frequently Asked Questions
ComfyUI Is it free to download from the official website?
Yes, it's free and open source: the code is hosted on GitHub under the GPL-3.0 license and runs on your own computer. The only three optional things that cost money are Comfy Cloud (the official paid cloud version), API / Partner nodes, and model files. See the "Is it free" section on this page.
Is there a Chinese version? How do I switch to it for free?
Yes. The core is free, and while the interface defaults to English, you can switch it to Chinese in Settings → Comfy → Locale—no need to download a separate "Chinese version installer."
ComfyUI Manager: what is it?
It's used to search for and install community custom nodes. Using nodes doesn't require writing code; when you need more features, use ComfyUI Manager to install community extensions.
Can FLUX be used for free?
Yes. FLUX.1 / FLUX.2 are both on the officially supported model list, the model files are free to download, and you can run them locally in ComfyUI as many times as you like. For beginner tutorials, see FLUX.1 and FLUX.2.
Where can I find ComfyUI basics tutorials?
This page is the first section of the basics tutorial: it explains nodes, workflows, and the official default text-to-image workflow. The next section covers installation: Installation: Windows Portable Package.
Next Section
Now that you have the concepts, install it on your computer: Installation: Windows Portable Package.