gnur000t/comfy-stack repository overview

⁠comfy-stack

It's ComfyUI pre-loaded with Tailscale and SwarmUI's custom nodes. All you need to bring is a persistent volume with models.

The general idea is that you run SwarmUI on your own computer and let Runpod (or any machine, really) do the inference. When you're done, terminate it, and all of your settings and output stay local. Next time, spin up another container/instance, attach a network volume with models and you're right where you left off.

⁠The Problem

• You're a friend of mine who regularly has to fit their whole life into like 4 suitcases.
• Setting up all the various frontends and connecting them together is a pain.
• Downloading models every time is, let's face it, wasteful.

⁠The Solution

• Install SwarmUI without a backend on your computer (or a cheap VPS)
• Install Tailscale
• Set up a Runpod persistent volume with models
• Deploy this image

⁠How to Use It

⁠Storage

You should mount a persistent volume at /workspace/ComfyUI/models, and from there use the normal folder structure for ComfyUI.

⁠Environment Variables

TAILSCALE_AUTHKEY - The auth-key to join the Tailnet. You should use ephemeral keys so that the machine is automatically removed when you're done. TAILSCALE_SERVER - The login server to use, if you're hosting your own, such as Headscale. If not specified, uses Tailscale's official SaaS platform TAILSCALE_HOSTNAME - The hostname to report to Tailscale. If not specified, uses comfy-<hostname> where <hostname> is the container's existing hostname

⁠Ports

If a single GPU is detected, ComfyUI will be located at port 8188. Additional GPUs will be given incrementing ports, for example, GPU 2 will get port 8189, and so on.