Setting up AI image generation on your own PC

What you need

• A PC with an NVIDIA graphics card. As a rough guide, one that runs recent open-world games reasonably well
• Disk space. A few GB for the app itself, plus about 6.5GB per model. Leave 50GB free
• Windows 10 / 11

Graphics card performance is measured as VRAM. 8GB is enough to start, but 12GB or more keeps resolution and batch size from becoming a problem. Open the Performance tab in Task Manager and look at the GPU section — the figure listed as “Dedicated GPU memory” is it.

Installing

There are several ways to install ComfyUI; this note covers two. Whichever you pick, the screen you get afterwards and the way you use it are much the same.

Desktop app (recommended): download the installer and run it. This is what the project itself recommends. Unless you have a reason not to, pick this one.

CLI: you type commands into a terminal. It suits you if:

• you want to control where it goes and which version you run
• you want to hand image generation work to a terminal-based AI such as Claude Code or Codex

The official CLI is built with coding agents in mind, so if that is your plan, start here.

Desktop app

1. Download the installer from comfy.org/download
2. Run it and follow the installer
3. Picking one at “Choose a starter workflow” downloads the models it needs automatically

CLI

Open the official page, comfy.org/cli. We are installing on our own PC from the command line, so pick Comfy CLI and Local ComfyUI.

Choosing the AI you use from the tabs below shows a prompt that finishes the setup once you paste it in. If you are handing the job to an AI, that is the quick way.

Here is the route where you type the commands yourself.

You need Python. If you do not have it, run this in a terminal.

winget install —id Python.Python.3.12 —exact

Install the CLI.

pip install comfy-cli

Set it up.

comfy setup

It is interactive, so answer the questions as they come.

Start it.

comfy launch

ComfyUI opens in your browser. From then on this one command starts it.

The commands do change, so if something does not work, check the official page.

The screen

ComfyUI is a screen where you wire boxes (nodes) together with lines to build up a pipeline.

The screen at startup

You may get an empty canvas, or a default workflow already on it. Either way, open Templates in the left menu and pick SDXL1.0: Text to Image (Simple).

The left menu

Icons run down the left edge of the screen. Click one and a panel opens beside it; click again and it closes.

The two you use first are Templates and Models.

The basic shape

The simplest text-to-image arrangement.

load the model --+-- positive prompt --+
                 |                     |
                 +-- negative prompt --+-- generate -> to image -> save
                 |                     |
                 +-- image size -------+

In actual node names:

Load Checkpoint has three outputs, going to KSampler, CLIP Text Encode (Prompt) and VAE Decode. This shape never changes, so leave the wiring alone at first and change only the text and the numbers.

Where the output goes

After VAE Decode you wire either Preview Image or Save Image. Preview Image just shows the result on the spot; Save Image writes it out to the output folder automatically. The template ships with Save Image.

Which one you want depends on how you handle what you generate. Preview Image if you save only the ones you like, by hand; Save Image if you keep everything and pick through it later. You can also wire both.

Preparing a model

ComfyUI on its own cannot produce anything. You need a model — a trained file of several GB. The art style comes from the model.

How to add one

Open a workflow whose models you do not have and an error appears at the top right of the screen. Press View details and the side panel lists the missing models with a download button.

The desktop app starts downloading right away, and progress shows on the download icon at the top right. The CLI install asks you where to put them.

If you already have models

In the desktop app, Menu → Desktop Settings → Storage → Add Shared Directory adds the whole folder. It needs a Relaunch afterwards.

To drop files in directly, the desktop app uses %LOCALAPPDATA%\Comfy-Desktop\ComfyUI-Shared\models\checkpoints and the CLI install uses ComfyUI\models\checkpoints. After adding a file, press r in the ComfyUI screen and it is picked up.

Which model to pick

If you plan to distribute or sell what you generate, checking the license is not optional. “May I redistribute the model” and “may I use the images commercially” are separate conditions, and conflating them is how people get burned.

If you are unsure, starting with Animagine XL 4.0 (anime) or RealVisXL V5.0 (photoreal) is a safe bet. The terms are permissive and both are easy to get results from.

Watch out for FLUX: despite the shared name, schnell allows commercial use and dev does not — the exact opposite of each other.

Generating an image

What to change first

• Pick your model under ckpt_name in Load Checkpoint
• Write the positive prompt in the upper CLIP Text Encode (Prompt)
• Write the negative prompt in the lower CLIP Text Encode (Prompt) (worst quality, low quality, bad anatomy is a reasonable start)
• Click Run

Anime models are trained on tags rather than sentences, so write the prompt as a comma-separated list like 1girl, blue hair, smile.

Settings to start from

The KSampler and Empty Latent Image values. The template ships with these.

Drop to something small like 512×512 and it falls apart immediately.

Summary

• If you get stuck, ask an AI 🫵
• Adult image generation is in the next note18+ 🤐