ComfyUI for Beginners: Make Your First AI Image, Step by Step
September 28, 2026 · Kova Core
In short: ComfyUI is a free, open-source app for creating AI images (and video and audio) on your own computer. Instead of one prompt box, you build a workflow from connected boxes called nodes. Install the Desktop app, download one model, press Run — then learn what each node does.
What is ComfyUI, and why use it?
Online image generators are easy, but they come with monthly fees, credits, queues, and your pictures stored on someone else's server. ComfyUI runs on your own machine:
- Free and open source — no subscription or credits.
- Private — your prompts and images stay on your computer.
- Flexible — works with many model families, from Stable Diffusion to newer models, plus video, upscaling and thousands of community add-ons.
- Reproducible — every image it saves remembers the exact workflow that made it.
The trade-off is a steeper learning curve, and you need a reasonably capable computer.
What computer do you need?
ComfyUI runs on Windows, macOS (Apple Silicon M-series) and Linux, and supports NVIDIA, AMD and Intel Arc graphics cards. An NVIDIA card is usually the smoothest experience.
The most important number is your graphics card's memory (VRAM). As a rough community rule of thumb:
| VRAM | What to expect |
|---|---|
| 4–6 GB | Stable Diffusion 1.5 works; newer models are slow or need low-memory versions |
| 8–12 GB | Comfortable with SDXL and optimised versions of many newer models |
| 16–24 GB or more | Most modern image models, and some video models, at good speed |
Apple Silicon Macs work too, using their shared memory — slower than a strong NVIDIA card, but fine for learning. Without any graphics card, ComfyUI can run on the CPU, but very slowly.
Step 1: Install ComfyUI
There are three official ways to install it:
- Comfy Desktop — a normal installer for Windows, macOS and Linux. Download it from comfy.org and follow the setup wizard. Best for beginners.
- Windows Portable — a zip file with everything included. Extract it and run. Popular with NVIDIA users who want the newest features.
- Manual install — for people comfortable with Python and Git; works with every operating system and graphics card.
Step 2: Download your first model
ComfyUI doesn't come with an AI model — you download model files (called checkpoints) separately. For a first test, ComfyUI's official guide uses Stable Diffusion 1.5 (v1-5-pruned-emaonly-fp16.safetensors), because it's small and runs on modest hardware.
Put checkpoint files in the models/checkpoints folder:
- Desktop:
ComfyUI/models/checkpointsinside your ComfyUI installation folder - Portable:
ComfyUI_windows_portable/ComfyUI/models/checkpoints
After adding a model, press R in ComfyUI to refresh the list.
Safety tip: download models only from trusted sources, such as the official model pages on Hugging Face, and prefer .safetensors files. Older .ckpt files from unknown sources can contain malicious code.
Step 3: Understand the default workflow
When ComfyUI opens, it usually loads the default text-to-image workflow. Here's what each box does, in the order the data flows:
| Node | What it does |
|---|---|
| Load Checkpoint | Loads the model. It outputs MODEL (the image generator), CLIP (understands your text) and VAE (converts to and from normal pixels) |
| CLIP Text Encode (positive) | Your prompt — what you want to see |
| CLIP Text Encode (negative) | What you don't want, e.g. "blurry, low quality, watermark" |
| Empty Latent Image | The blank canvas: width, height and how many images |
| KSampler | The engine. It starts from random noise and removes it step by step, guided by your prompts |
| VAE Decode | Turns the finished result into a normal image |
| Save Image | Shows the picture and saves it to ComfyUI's output folder |
The wires carry data from one node's output to the next node's input — the colours show which types can connect.
Step 4: Write a prompt and press Run
In the positive prompt box, describe the picture. Be specific about subject, setting, style and lighting, for example:
"a cosy coffee shop on a rainy evening, warm lights, people reading, watercolour illustration, highly detailed"
Then click Run, or press Ctrl + Enter. The first run takes longer while the model loads. Your image appears in the Save Image node and is saved in the output folder.
The KSampler settings that matter
| Setting | What it means | Good starting point |
|---|---|---|
| seed | The random starting noise. Same seed + same settings = the same image | Leave random; fix it when you like a result |
| control after generate | Whether the seed changes after each run | "randomize" to explore, "fixed" to fine-tune |
| steps | How many clean-up passes the sampler makes | 20–30 |
| cfg | How strictly to follow your prompt | 6–8 for SD 1.5 and SDXL; around 1 for distilled models like Flux, which use a separate guidance setting |
| sampler_name | The algorithm that removes the noise | euler or dpmpp_2m |
| scheduler | How quickly noise is removed over the steps | normal or karras |
| denoise | How much to change the starting image | 1.0 for text-to-image |
5 tips that save beginners hours
- Drag an image in to load its workflow. Every image ComfyUI saves contains the workflow that made it. Drop a ComfyUI-made PNG onto the canvas and the whole setup appears — the fastest way to learn from others.
- Change one setting at a time with a fixed seed, so you can see exactly what each setting does.
- Match the image size to the model — about 512×512 for SD 1.5, and about 1024×1024 for SDXL and most newer models. The wrong size often gives stretched or doubled subjects.
- Start from templates. Recent versions include a Templates browser with ready-made workflows for image-to-image, upscaling, inpainting and video. Starting from a working example beats building from scratch.
- Be careful with custom nodes. The community has built thousands of extensions, usually installed through the ComfyUI-Manager extension. They're powerful, but they run code on your PC — stick to popular, well-maintained ones.
After your first image
- Shrink your results for sharing or websites with our Image Compressor, or crop them with the Image Editor.
- Know what your images carry. Every PNG that ComfyUI saves contains your full workflow and prompt. That's handy for you, but if you'd rather not share them, strip them with our Photo Location Remover before posting.
- Check the licence of each model — some don't allow commercial use.
- Be responsible: don't create images of real people without their consent, or anything illegal or harmful.
New to AI in general? Start with The Generations of AI Explained, or run a private AI chatbot next to ComfyUI with How to Run AI on Your Own Computer.