ComfyUI is the most powerful free tool for generating AI images on your own computer — and the one beginners bounce off most often, because its website does not look like normal software at all. There is no big friendly installer on the front page, and the program itself, a canvas of connected boxes, looks like a circuit diagram. None of that matters. Installing ComfyUI on Windows the portable way takes about fifteen minutes, needs no programming, and this guide walks through it in order: what your PC needs, the download, your first image model, your first picture, and the four errors almost everyone meets on day one (all four have simple fixes).
What ComfyUI Is, in One Paragraph
ComfyUI is a free, open-source program that runs AI image models — Stable Diffusion, SDXL, FLUX and others — entirely on your own machine. Instead of typing a prompt into a website, you arrange small boxes ("nodes") on a canvas: one loads the model, one holds your prompt, one does the generating, one saves the picture. The boxes look technical, but the default arrangement is already built for you, and for your first weeks you will change nothing except the prompt. Everything runs locally: no subscription, no per-image cost, no uploading your ideas to anyone's server. The project lives on GitHub as Comfy-Org/ComfyUI, where it is one of the most popular AI tools in the world.
Before You Start: What Your PC Needs
AI image generation is heavy work, and honesty here saves frustration later. What matters most is your graphics card (GPU) and its memory, called VRAM:
- NVIDIA graphics card with 8GB VRAM or more — the comfortable zone. SDXL and modern models run well.
- NVIDIA card with 4–6GB VRAM — workable. Start with the smaller SD 1.5 models and use the low-memory setting described below.
- AMD graphics card — possible on Windows, but the portable build below is made for NVIDIA. AMD owners should use the official Desktop edition from comfy.org instead, which handles the setup for them.
- No graphics card (integrated graphics only) — ComfyUI can run on the CPU alone, and this guide includes that option, but expect minutes per image instead of seconds. It is a way to learn the tool, not to love it.
- Disk space — reserve at least 20GB. The program itself is a few gigabytes, and each image model you download is 2–7GB more.
- Windows 10 or 11, and a program that can open .7z archives — the free 7-Zip handles it.
Step 1: Download the Portable Build
The portable build is a self-contained folder: Python and every dependency are already inside it, so nothing gets installed into Windows itself and deleting the folder removes everything cleanly. That is why it suits beginners.
- Open the ComfyUI project page and click Releases in the right-hand column.
- In the latest release, look under Assets and download the Windows portable package — the file named like ComfyUI_windows_portable_nvidia.7z. (A "cpu" variant exists too, for machines with no NVIDIA card.)
- The download is several gigabytes; let it finish completely before continuing.
Step 2: Extract It to a Simple Folder
- Right-click the downloaded .7z file and extract it with 7-Zip (7-Zip → Extract here).
- Move the resulting ComfyUI_windows_portable folder somewhere simple, like C:\ComfyUI or a second drive with space. Avoid putting it inside cloud-synced folders (OneDrive, Dropbox) — syncing thousands of small files causes strange errors.
- Keep the folder path plain: short, and without unusual characters.
Step 3: Start ComfyUI for the First Time
- Open the portable folder and double-click run_nvidia_gpu.bat (or run_cpu.bat if you have no NVIDIA card).
- A black window opens and fills with text. This is normal — it is the program starting its local server. The first start can take a few minutes.
- When it is ready, your browser opens by itself at http://127.0.0.1:8188 — that address is just your own computer talking to itself; nothing is online. If the browser does not open, type that address in manually.
- You will see the node canvas with a ready-made starter workflow already loaded. Leave the black window open while you work — closing it closes ComfyUI.
Step 4: Add Your First Image Model
ComfyUI is the engine; a model (also called a checkpoint) is the brain it thinks with, and the portable build ships without one. To add one:
- Download a checkpoint file (it ends in .safetensors). For a first model, choose a Stable Diffusion 1.5 checkpoint (about 2–4GB) if your VRAM is limited, or an SDXL checkpoint (about 6–7GB) if you have 8GB VRAM or more. Model libraries such as Hugging Face and Civitai list thousands, most free.
- Copy the file into ComfyUI\models\checkpoints inside your portable folder.
- Back in the browser, refresh the page. In the first box on the canvas (Load Checkpoint), click the model name field and select your file.
Step 5: Generate Your First Image
- Find the two text boxes: the upper one is the positive prompt (what you want), the lower one the negative prompt (what you want to avoid). Type something simple into the positive box, like "a small cottage by a lake at sunset, detailed photograph".
- Click Queue Prompt at the top-right.
- Watch the boxes light up in turn as the work flows through them. In under a minute (seconds, on a strong GPU), your image appears in the Save Image box at the end.
- Finished images are also saved as files in ComfyUI\output, so nothing is lost when you close the program.
Congratulations — everything from here is experimentation: new prompts, new models, and ready-made workflows other users share as simple drag-and-drop files.
Two Settings Worth Knowing Early
You can ignore almost every number on the canvas except two. The image size (in the Empty Latent Image box) controls both quality and memory use: 512×512 is the natural size for SD 1.5 models, 1024×1024 for SDXL. Asking a small model for a huge picture is the most common cause of strange, doubled images — match the size to the model family. The seed (in the KSampler box) is the recipe number for randomness: leave it on "randomize" while exploring, then copy a seed you liked if you want to reproduce an image and vary it slightly. Those two settings, plus your prompt, are ninety percent of everyday ComfyUI.
The Four Errors Everyone Meets (and Their Fixes)
- "CUDA out of memory" — your model is bigger than your VRAM. Right-click run_nvidia_gpu.bat, choose Edit, and add --lowvram to the end of the command inside it, then save and start again. Low-VRAM mode is slower but lets modest cards run big models. Lowering the image size (for example 512×512) helps too.
- Images come out completely black — usually a missing or mismatched VAE (the part that turns data into a visible picture). Most modern checkpoints have one built in; if yours does not, download the matching VAE file into models\vae and select it in the workflow's VAE box.
- "Torch not compiled with CUDA enabled" — you started the wrong launcher. Close everything and use run_nvidia_gpu.bat, not the CPU one.
- Red boxes / "missing custom nodes" — you have loaded a workflow someone shared, and it uses add-on nodes your copy does not have yet. This is normal and fixable later with the popular ComfyUI-Manager add-on; for now, the default workflow needs no add-ons at all.
Frequently Asked Questions
Is ComfyUI really free? Yes — the program and most models are free and open source, and because everything runs on your PC there is no subscription and no per-image charge. Your only cost is electricity.
ComfyUI or the simpler Stable Diffusion WebUI — which should a beginner choose? The WebUI (AUTOMATIC1111) looks friendlier on day one; ComfyUI is faster, lighter on VRAM, and where most new models arrive first. Many people learn ComfyUI second and stay with it.
Does ComfyUI work on Mac or Linux? Yes — this guide covers Windows because the portable build is Windows-only, but Mac and Linux installs are covered in the project's README, and the interface is identical once running.
How much disk space will this take over time? Models add up quickly — five checkpoints can pass 25GB. Keep them in the one checkpoints folder so tidying up stays easy.
Where can I find more models and workflows? Start from the AI image generation category here for the tools, and see our roundup of the best free AI tools that run locally for what else your PC can now do. When you outgrow the defaults, the ComfyUI project page links to examples and documentation.
That is the whole journey: one download, one folder, one double-click, and the most capable free AI image studio in the world is running on your own machine — no account, no subscription, and no code written by you at any point.