ComfyUI Beginner's Guide: Install It and Make Your First Image

What ComfyUI is

ComfyUI appeared in early 2023 and is now maintained by Comfy Org under the GPL-3.0 license. Where most image tools give you a text box and a button, ComfyUI shows the pipeline itself as a graph of nodes: load a model, encode the prompt, denoise, decode to pixels, save. You connect nodes with wires. The finished graph is a workflow, a JSON file you can save, share and rerun exactly.

That design has a cost, the learning curve, and a payoff: new model families are almost always supported in ComfyUI first, and anything you build is reproducible. As of September 20, 2026, it is the default choice for anyone who wants to go past one-click generation. The source is at github.com/Comfy-Org/ComfyUI and the manual at docs.comfy.org.

Hardware: how much VRAM each model family needs

As with language models, GPU memory is the deciding resource. The difference is the failure mode: an image model that overflows VRAM does not get a little slower, it gets ten times slower or stops with an out-of-memory error.

Model family	Comfortable VRAM	Cards from our hardware database that fit
Stable Diffusion 1.5	4–6 GB	GTX 1660, RTX 2060, RTX 3050 6 GB and up
SDXL and derivatives	8 GB	RTX 3060, RTX 4060, RTX 5060, RX 7700 XT
Flux.1 in FP8, recent 6B–12B image models	12–16 GB	RTX 3060 12 GB, RTX 4070, RTX 5070 Ti, RX 9070 XT
Flux.1 at full precision	24 GB	RTX 3090, RTX 4090, RTX 5090, RX 7900 XTX
Video (Wan family and similar)	16–24 GB and up	RTX 4090, RTX 5090; lower with quantized builds, slowly
Apple Silicon	16 GB unified memory minimum	Any M-series chip, typically 2–4× slower than a comparable RTX card

Orders of magnitude for a 1024 × 1024 image. Card list from the BestLLMfor hardware database, September 20, 2026.

NVIDIA is the frictionless path. AMD cards work well on Linux through ROCm; Windows support is improving but still takes more effort. Budget disk space too: a single model is 2 to 25 GB, and collections grow fast. If you are unsure what your card has, see what VRAM is and how to check it.

Three ways to install ComfyUI

Route	Best for	What to know
ComfyUI Desktop	Beginners on Windows (NVIDIA) or Apple Silicon Macs	A normal installer from comfy.org. Sets up an isolated Python environment and updates itself.
Windows portable	Windows users who want everything in one folder	Download the archive from the GitHub releases page, extract it, double-click run_nvidia_gpu.bat. Nothing is installed system-wide.
Manual install	Linux, AMD cards, people who want control of the environment	Git clone, a Python virtual environment, the PyTorch build for your GPU, then the project requirements.

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI
python -m venv venv && source venv/bin/activate

# PyTorch: copy the exact command for your GPU from pytorch.org
pip install torch torchvision torchaudio

pip install -r requirements.txt
python main.py

The terminal prints the local address to open. Two launch flags are worth knowing: --lowvram for cards with 6 GB or less, and --listen to reach the interface from another device on your network.

Where model files go

This is the first place people get stuck: the interface opens, you press Run, and an error names a missing file. ComfyUI includes no models, and each file type has its own folder.

Folder	What goes there
models/checkpoints	All-in-one models such as SDXL: one .safetensors file holds the diffusion model, text encoder and VAE. Simplest to start with.
models/diffusion_models	The diffusion model alone, for recent families shipped as separate parts (Flux, Wan and others).
models/text_encoders	The encoder that turns your prompt into vectors. Often the second-largest file.
models/vae	The decoder that turns the computed result into a visible image.
models/loras	LoRAs: small files that add a style, character or concept to a base model.

The shortcut that saves an hour: open Workflow → Browse Templates. Pick an official template and ComfyUI detects the missing files and offers to download each into the right folder. Start there instead of hunting for files by hand. When you do download manually, most models live on Hugging Face; our guide to Hugging Face explains how to read a repository.

Your first image, step by step

1. Load an official template. Choose the basic image-generation template that matches your VRAM: SDXL from 8 GB, a recent FP8 model from 12 GB.
2. Accept the downloads and watch the terminal until they finish. An interrupted download leaves a truncated file that causes obscure errors later.
3. Write the prompt. The text-encode node wired to the positive input holds your description; the one wired to negative lists what you do not want.
4. Press Run. Nodes light up in sequence. The first generation is slow while the model loads into VRAM; the next ones are much faster.
5. Find the image in the Save Image node and in the output folder.

Three sampler settings are enough to begin. Seed fixes the image: same seed and settings, same result. Steps trades time for refinement, 20–30 for a standard model, 4–8 for a "turbo" or distilled one. CFG sets how strictly the image follows the prompt; too high and it oversaturates and distorts.

Workflows: how to learn from other people's images

Because ComfyUI embeds the workflow in every PNG it saves, the fastest way to learn is to drag someone else's output into the window. The full graph reloads with prompt, seed and settings intact. This is why the community shares images rather than tutorials.

If a loaded workflow shows red nodes, it depends on custom nodes you do not have. ComfyUI-Manager, included in recent installs, lists and installs what is missing. Stay lean: each extension adds Python dependencies, and an install carrying fifty of them tends to break on the next update.

A workflow runs code. Custom nodes are Python programs executing with your permissions. Install only known, maintained extensions through the manager, and be suspicious of a workflow that demands an extension you cannot find anywhere else.

Common first-week errors

Symptom	Usual cause	Fix
CUDA out of memory	Model or resolution exceeds VRAM	Lower the resolution, use an FP8 or quantized build, or launch with --lowvram
Black image or colored noise	VAE or text encoder does not match the model	Go back to the official template for that model family
Generation takes minutes	Running on CPU	Check the startup log for your GPU; reinstall the PyTorch build that matches your card
Red nodes on load	Missing custom nodes	Install them through ComfyUI-Manager, then restart

Licenses: what you may do with the images

ComfyUI is free software, but every model carries its own license.

Model	License	Commercial use of outputs
Flux.1 schnell	Apache 2.0	Allowed
Flux.1 dev	Black Forest Labs non-commercial license	Restricted: read the terms; a paid license exists
SDXL	OpenRAIL++	Allowed, with use restrictions listed in the license

Check the license of the exact model and of any LoRA you add before using an image for a client or a product.

Is ComfyUI the right tool for you?

If all you want is to type a prompt and get a picture, a one-button interface will serve you better on day one. ComfyUI becomes the right choice the moment you want to reproduce a result exactly, chain several stages (generate, upscale, retouch), or use a model released last week. For a complete path with settings per GPU, see the Local Image AI Kit. Hardware figures on this page are open through the BestLLMfor public API (CC BY 4.0) and our MCP server.