ComfyUI on Edge, your workflows on your servers

Deploy

ComfyUI on CPU in five steps

A CPU build of PyTorch, models from your bucket, and a proxy with authentication in front. ComfyUI has none of its own.

~/comfy — zsh

$ edge compute create --name comfy \

--size s-8vcpu-16gb --disk 160 \

--image ubuntu-24 --region london \

--script comfy-setup

✓ vm-6b02e9 running · 64.34.86.90

$ edge storage sync comfy-models/ models/

✓ SD 1.5 checkpoint + 3 LoRAs synced

$ python main.py --cpu --listen 127.0.0.1

✓ Starting server · 127.0.0.1:8188

✓ Live at https://comfy.acme.com

Per-image fees$0

comfy.acme.comRunning

RuntimeComfyUI · --cpu

CheckpointSD 1.5 · 512 px

1. 01

   Write a bootstrap script

   Installs ComfyUI with CPU-only PyTorch wheels, Nginx for auth, and the Edge CLI for model syncs.

   comfy-setup.shCopy

   #!/bin/bash
   set -e
   apt-get update
   apt-get install -y python3-venv git nginx apache2-utils
   curl -fsSL https://edge.network/install.sh | sh
   git clone https://github.com/comfyanonymous/ComfyUI /opt/comfy
   cd /opt/comfy && python3 -m venv .venv && . .venv/bin/activate
   pip install torch torchvision torchaudio \
     --index-url https://download.pytorch.org/whl/cpu
   pip install -r requirements.txt

2. 02

   Create the VM and the buckets

   16 GiB of memory is the practical floor for diffusion on CPU. --disk 160 leaves room for a few checkpoints.

   shellCopy

   $ edge compute scripts create --name comfy-setup --file comfy-setup.sh
   $ edge compute create --name comfy --size s-8vcpu-16gb --disk 160 \
       --image ubuntu-24 --region london --script comfy-setup
   $ edge storage create comfy-models
   $ edge storage create comfy-outputs

3. 03

   Run it as a service

   Models sync from the bucket before each start, so every VM has the same set. ComfyUI stays on localhost.

   /etc/systemd/system/comfy.serviceCopy

   [Service]
   WorkingDirectory=/opt/comfy
   EnvironmentFile=/etc/comfy.env   # EDGE_API_KEY
   ExecStartPre=/usr/local/bin/edge storage sync \
     comfy-models/ /opt/comfy/models/
   ExecStart=/opt/comfy/.venv/bin/python main.py --cpu \
     --listen 127.0.0.1 --port 8188
   Restart=always
   
   [Install]
   WantedBy=multi-user.target

4. 04

   Put basic auth in front

   Create the password file with htpasswd. The WebSocket headers keep the editor's live progress working.

   /etc/nginx/sites-enabled/comfyCopy

   server {
     listen 443 ssl;
     server_name comfy.acme.com;
     ssl_certificate     /etc/ssl/origin.pem;   # origin certificate
     ssl_certificate_key /etc/ssl/origin.key;
   
     auth_basic "ComfyUI";
     auth_basic_user_file /etc/nginx/.htpasswd;
   
     location / {
       proxy_pass http://127.0.0.1:8188;
       proxy_http_version 1.1;
       proxy_set_header Upgrade $http_upgrade;
       proxy_set_header Connection "upgrade";
     }
   }

5. 05

   Add the CDN and queue a workflow

   One domain for the editor and API, one serving outputs from the bucket with image transforms on.

   shellCopy

   $ edge cdn create --name comfy
   $ edge cdn domains add cdn-a1b2c3 \
       --domain comfy.acme.com --origin https://<vm-ip>
   $ edge cdn domains add cdn-a1b2c3 --domain img.acme.com \
       --origin-type storage --storage-bucket comfy-outputs --jit-image
   
   # payload.json = { "prompt": <workflow exported in API format> }
   $ curl -u artist:$PASS https://comfy.acme.com/prompt \
       -H "Content-Type: application/json" -d @payload.json

Prefer to hand it off? Give the job to your AI agent or have our engineers do it.