So the first thing you do is a git clone Then cd ComfyUI/custom_nodes. Then issue git clone Last thing to do, navigate to ComfyUI again and issue pip install -r requirements.txt. So I’m skipping some of the PyTorch stuff and that’s important and all, but this is just a real condensed summary that you’re going to have to fill the gaps of. Sorry. I might be too stoned to do a proper doc on it, but it’ll get you pointed in the right direction.

So once all the requirements are set you just run python Boom, you’re into it. You should have a URL in the window it’s running in that you can click, so click it, and you’ll be greeted by the default ComfyUI web interface.

Hopefully you have the manager installed, so you should see the manager button to the right. Click it. Use that sumbitch to install all of your plugins and models and it’ll save you a shitload of heartburn. I got up and running with ControlNets and SDXL models real fast. See, the key is, you have to match ControlNet model with stable diffusion model. For the SDXL models you’ll need SDXL compatible controlnet models, which are all labeled “controlnet-lora” in the manager. Boom, snag those. Snag all the controlnet models you intend to use. Just make sure you match like for like. It’ll give you an error if the controlnet model you selected isn’t compatible with the base model you’re trying to use.

Man, this stuff is really just a fun lockpick to dick with. But so far, ComfyUI has generated the absolute coolest results. You can also grab extra models from CivitAI as well as ComfyUI workflows. Unfortunately you’re going to have to pick up a lot of information from YouTube, though I can vouch for a number of the videos I’ve seen so far not being horribly insufferable. They actually do a good job mostly of laying out the visual structures of workflows and why things will or will not work. It’s been enlightening to watch these videos and compare them with my own results. I like it.