6 min read

ComfyUI on your own server: setup, models and workflows

How to work with ComfyUI on a rented GPU server: the model folder layout, the node manager, saving workflows and reaching the interface.

ComfyUI on your own server is a persistent workspace: models live on the disk, workflows are saved as files, and you start the server only when you need it.

Folder layout

models/checkpoints — main SDXL, FLUX and other checkpoints.
models/loras — LoRA files.
models/vae, models/clip, models/unet — components that newer models such as FLUX and Wan use.
user/default/workflows — saved workflows.

Node manager

Install nodes through the manager inside the interface, and only what your workflow needs. The more nodes, the higher the risk of dependency conflicts; if ComfyUI will not start after an install, roll back the last change.

Access and data

Open the interface from the server card. Upload files with SFTP or wget, and download results before you delete the server: the disk is deleted with it.

# inside the ComfyUI folder
wget -O models/checkpoints/model.safetensors "<link>"

Saving money

Keep the server stopped when you are not working: GPU minutes are not billed. The model disk is billed the whole time, so do not keep extra weights.