ComfyUI cloud GPU Europe: install it and access it safely

This guide takes you from nothing to a working ComfyUI cloud GPU Europe setup: a European host found on Cloudj, ComfyUI installed by you over SSH, models stored on the instance disk, and the web interface opened safely through an SSH tunnel instead of the open internet. It suits illustrators, studios and developers in Europe who want ComfyUI on a fast card without buying one, or anyone who prefers European hosts for their image work.

For choosing European countries in general, read the GPU cloud Europe guide.

Step 1: choose the card

ComfyUI runs on almost any modern NVIDIA GPU, but VRAM decides which models and resolutions are comfortable. As general guidance:

• RTX 4090, 24 GB: SDXL with LoRAs and ControlNet, most popular community workflows, and many newer image models with sensible settings.
• RTX 5090, 32 GB: more headroom for large image models and video workflows. It is a Blackwell card, so use a recent PyTorch build.
• L40S, 48 GB class: large models at higher precision, heavy video workflows, several models loaded at once. It needs a higher account level.

If you are unsure, start with a 24 GB card and move up only if you hit out-of-memory errors.

Step 2: search European countries

GPU location on Cloudj is a country filter over a live list of host machines. Open the RTX 4090 deploy page and enter:

DE, NL, FR

Results show the country (never a city), reliability, max duration, download speed and bandwidth price. Download speed matters more than usual here: you will pull several gigabytes of checkpoints. If a machine is listed, you can deploy it. If not, stock changes during the day; add more European codes (BE, PL, CZ, SE) or search by a 24 GB minimum VRAM instead of a model name. Live "from" prices are on GPU cloud.

Step 3: deploy with the right disk

Choose the Ubuntu 22.04 + CUDA (SSH) template and select your SSH key. Then set the disk size carefully, because it cannot be changed after deploy. A typical ComfyUI disk holds:

• the Python environment and ComfyUI itself (a few GB);
• one or two base checkpoints (several GB each for SDXL-class models, more for larger ones);
• LoRAs, ControlNets, VAEs and upscalers;
• your outputs.

Err on the generous side. The GPU disk size guide has a worksheet for estimating it.

Step 4: connect and install ComfyUI

Open the instance's Connection tab for the host and port, then connect. A new key can take up to three minutes to be accepted after the instance starts.

ssh -p <port> root@<host>
nvidia-smi

Install ComfyUI in a virtual environment. Cloudj does not preinstall it; this is your own setup:

apt-get update && apt-get install -y git python3-venv
cd /workspace
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
python3 -m venv venv && . venv/bin/activate
pip install torch torchvision torchaudio
pip install -r requirements.txt

If you rented a Blackwell card, install a PyTorch build for a recent CUDA version. Then check the GPU is visible:

python -c "import torch; print(torch.cuda.get_device_name(0))"

Step 5: put models on the disk

Download checkpoints directly on the instance rather than uploading them from your computer. It is faster and avoids paying upload time twice. ComfyUI expects models in folders such as models/checkpoints, models/loras and models/vae:

cd /workspace/ComfyUI/models/checkpoints
wget -c "<direct download URL of the checkpoint>"

For gated models, log in with the Hugging Face CLI on the instance first. Respect each model's licence, especially for client or commercial work.

Step 6: start ComfyUI and reach it safely

Start ComfyUI listening on localhost only:

cd /workspace/ComfyUI && . venv/bin/activate
python main.py --listen 127.0.0.1 --port 8188

Do not expose port 8188 to the internet. ComfyUI has no built-in authentication, and an open instance can be used by anyone who finds it. Instead, forward the port from your computer over SSH:

ssh -p <port> -L 8188:localhost:8188 root@<host>

Leave that session open and visit http://localhost:8188 in your browser. The interface behaves as if it were local, but all the work happens on the European GPU. Run ComfyUI inside tmux so it keeps going if your SSH connection drops. The SSH into a GPU instance article covers keys, config files and tunnels in more detail.

Step 7: save your work and shut down

• Copy outputs and any custom workflows back regularly:

rsync -avP -e "ssh -p <port>" root@<host>:/workspace/ComfyUI/output/ ./comfy-output/

• Remember each machine shows "available until" a date; after it, the instance and disk are removed.
• Stop keeps your disk and charges storage only, but releases the GPU. When you start again you may wait until that GPU is free.
• Destroy deletes everything and stops charges. Use it when the project is done and backed up.

If you run the same setup often, consider a public custom Docker image with ComfyUI preinstalled (it needs a higher account level), or a setup script you paste after each deploy. The Stable Diffusion GPU in Japan article shows the same workflow for a host in Asia.

FAQ

Does Cloudj have ComfyUI preinstalled?

No. You install ComfyUI yourself on the Ubuntu + CUDA template, or use your own public Docker image.

Can I open ComfyUI with a public link?

You could, but you should not. ComfyUI has no login. Use SSH port forwarding, or put it behind a proxy with authentication.

Which European country is best for ComfyUI?

For a solo artist, any nearby host works. Search DE, NL, FR and pick on price, reliability and download speed.

What if the European host disappears?

Deploy a new one, rerun your setup script and restore models and workflows from your backup.

Start creating

Open the RTX 4090 deploy page, enter DE, NL, FR and see which machines are listed now. Compare live prices on GPU cloud.