A practical guide for designers — no jargon, no skipped steps, no assumed knowledge. Pick the path that matches your machine. By the end of the chosen section, you'll be staring at a working node graph generating real images.
Four ways to run ComfyUI. There's no "right" one — pick the one that matches your hardware. If you have no GPU and no patience, skip straight to Google Colab.
The official ComfyUI Windows portable package — no Python setup, no environment hassle. Download, unzip, double-click. This is the fastest path.
Press Win + R, type dxdiag, press Enter. Click the "Display" tab. Read what's next to "Name". You're looking for an NVIDIA GeForce card.
Open github.com/comfyanonymous/ComfyUI/releases in a browser. Under the latest release, find the file ending in _nvidia.7z (it's around 2 GB).
Look for something like ComfyUI_windows_portable_nvidia.7z. Avoid any file with "directml" or "cpu" in the name — those are for AMD or CPU-only setups, which are slow.
You'll need 7-Zip to extract the file. Download it from 7-zip.org if you don't have it. Right-click the downloaded archive → 7-Zip → Extract Here.
This creates a folder called ComfyUI_windows_portable. Important: put this folder somewhere with at least 40 GB of free space — your C:\ drive root, or a fast external SSD.
Inside the extracted folder you'll see two batch files:
run_nvidia_gpu.bat — use this onerun_cpu.bat — only if you don't have NVIDIADouble-click run_nvidia_gpu.bat. A black command window opens. After about 30 seconds your default browser opens to 127.0.0.1:8188 with the ComfyUI interface.
If the browser opened and you see a node graph with a few connected boxes, ComfyUI is installed. You can try clicking Queue Prompt right now — but you'll get an error because no model is loaded yet. That's next.
ComfyUI by itself can't generate anything — it needs a model (also called a "checkpoint"). For this workshop, start with one of these:
Place the downloaded .safetensors file inside: ComfyUI_windows_portable\ComfyUI\models\checkpoints\
Back in the ComfyUI browser tab, click Refresh (top right). In the node called "Load Checkpoint", click the dropdown — your model should appear. Select it. Click Queue Prompt. Wait 10–30 seconds. You'll see your first AI image.
If you see an image on the right side of the screen, your installation works. You're ready for Workshop Session 1.
M1, M2, M3, M4 Macs run ComfyUI well using Apple's Metal acceleration. The install needs Terminal but the steps are copy-paste.
Open Terminal (press Cmd + Space, type Terminal, hit Enter). Paste this entire line and press Enter:
It'll ask for your Mac password. Type it (you won't see characters, that's normal). Wait 5–10 minutes. When it finishes, run the two lines it tells you to run (about adding Homebrew to your PATH).
That last line takes about 5 minutes. Don't worry about the long output.
Open that URL in Safari or Chrome. Same interface as Windows.
Every new session, open Terminal and run: cd ~/ComfyUI && source venv/bin/activate && python main.py. Or save those three lines as a shell script you double-click.
Same as the Windows section — place .safetensors files inside ~/ComfyUI/models/checkpoints/.
Apple Silicon is good, not great. Generation is roughly 3-5× slower than a similarly priced NVIDIA card. For Flux models specifically, you'll want a 16 GB or 24 GB unified memory Mac. If you have an 8 GB M-series, stick with SD 1.5 or SDXL Turbo.
If you're on Linux you probably know what you're doing. Here's the short version for Ubuntu 22.04+ with an NVIDIA GPU.
nvidia-smi should return your GPUThe --listen 0.0.0.0 flag exposes ComfyUI on your network — useful for headless Linux boxes. Access it from another machine at http://YOUR-LINUX-IP:8188. If using only locally, omit the flag.
Run ComfyUI on a free Google GPU from any laptop — even a 10-year-old one. This is the path for anyone without an NVIDIA card. Setup is genuinely two clicks.
We've prepared a Colab notebook with ComfyUI and a base model pre-configured. Click the link below to open it.
Click the ▶ play button on each cell from top to bottom. The first cell installs ComfyUI (~2 min), the second downloads a model (~2 min), the third launches the server.
At the end, you'll see a URL like https://random-words.trycloudflare.com. Click it. ComfyUI opens in a new tab, running on Google's GPU.
You get a free NVIDIA T4 (better than most laptops), no local install, and you can save your workflows to Google Drive. The catch: free tier disconnects after ~4 hours of inactivity. Plenty for any single workshop session.
Free Colab is metered. Heavy daily use will eventually rate-limit you. For ₹800/month Colab Pro gives you longer sessions and better GPUs. Worth it if you start using ComfyUI seriously.
Three things to set up before you do anything else. Together they take 10 minutes and unlock 80% of what makes ComfyUI useful for architecture.
The most important add-on. Lets you install everything else from within the UI instead of via command line.
Install method:
ComfyUI/custom_nodes/ foldergit clone https://github.com/ltdrdata/ComfyUI-Manager.gitControlNet is what makes ComfyUI useful for architects — it lets you feed in your linework, depth maps, or CAD exports. Install via the Manager:
controlnetLoRAs are small style add-ons (typically 50-300 MB) that fine-tune the model toward a specific aesthetic. For architecture, two are worth starting with:
Search for "architecture" on civitai.com. Download to ComfyUI/models/loras/.
If you've ticked every box, you're set. See you in the first session. If anything broke, the contact page has WhatsApp — fastest channel for setup problems.
Now that ComfyUI runs, the next skill is writing prompts that actually produce architecture rather than generic AI mush. That's a separate, longer guide.