*[DediGPU](https://dedigpu.com/) — Markdown mirror of [https://dedigpu.com/guides/run-comfyui-on-a-rented-gpu-server](https://dedigpu.com/guides/run-comfyui-on-a-rented-gpu-server) · updated 2026-09-02 · index for LLMs: [llms.txt](https://dedigpu.com/llms.txt) · everything: [llms-full.txt](https://dedigpu.com/llms-full.txt)*

# Run ComfyUI on a rented GPU server: setup, checkpoints, secure access

> Order a server with the ComfyUI template, reach it safely through an SSH tunnel, bring your checkpoints, add custom nodes, and keep it running for months.

*Image & video · published 18 July 2026 · updated 2 September 2026 · by DediGPU engineering*

In short:

- Order an RTX 5090 or 4090 with the ComfyUI template; it listens on port 8188 with no login, so reach it through an SSH tunnel.
- Checkpoints are not bundled: pull them from Hugging Face onto the local NVMe; egress is free.
- ComfyUI runs as a systemd service and survives reboots; sync outputs out on a cron, there are no backups.

ComfyUI on a rented card is the cheapest way to run Flux, SDXL and the video models without buying a workstation, and the template does the install. What remains is reaching it safely and getting your models onto it.

## 1. Order

An [RTX 5090](https://dedigpu.com/gpu/rtx-5090) (32 GB, $163) is the sweet spot for Flux and video; an [RTX 4090](https://dedigpu.com/gpu/rtx-4090) ($139) is enough for SDXL and Flux at FP8. Pick Ubuntu 24.04 and the **ComfyUI** template. It installs ComfyUI, ComfyUI-Manager and the Flux and SDXL loaders, on port 8188, with no checkpoints bundled.

## 2. Reach it through a tunnel, not the open port

The port is public at first boot and ComfyUI has no login. Do not expose it. Tunnel instead:

```
ssh -N -L 8188:localhost:8188 root@203.0.113.10
# then open http://localhost:8188 on your machine
```

```
# and close the port to the world
ufw default deny incoming && ufw allow 22 && ufw enable
```

For a shared team, put Caddy in front with basic auth and TLS on a hostname you own; egress is free, so the traffic costs nothing.

## 3. Bring your checkpoints

```
cd /opt/dedigpu/comfyui/models
# Hugging Face, resumable
pip install -U huggingface_hub
huggingface-cli download black-forest-labs/FLUX.1-dev --local-dir checkpoints/flux1-dev --include "*.safetensors"
# or from your machine
scp my-model.safetensors root@203.0.113.10:/opt/dedigpu/comfyui/models/checkpoints/
```

The NVMe is local and fast (400 GB on the 5090 server); a Flux checkpoint loads in a couple of seconds after the first run.

## 4. Custom nodes

ComfyUI-Manager is installed: *Manager → Install custom nodes* in the UI, or on the shell:

```
cd /opt/dedigpu/comfyui/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
systemctl restart comfyui
```

## 5. Keep it running

ComfyUI runs as a systemd service and comes back after a reboot. Logs and control:

```
systemctl status comfyui
journalctl -u comfyui -f
nano /opt/dedigpu/comfyui/env      # extra arguments, e.g. --fp8_e4m3fn-unet --highvram
```

Outputs land in `/opt/dedigpu/comfyui/output`; sync them out with `rclone` or `rsync` on a cron, because the server keeps no backups.

## 6. What it costs

The 5090 server is $163 a month, $146.70 on a 12-month term; there is no per-image cost, no egress and no idle charge. At a thousand Flux images a day that is under a cent per image with the card busy a third of the time.