ComfyUI on SwissGPU

Introduction

ComfyUI is a node-based interface for running diffusion and image generation workflows. On a SwissGPU server it can be installed directly over SSH, started on port8188, load your own models, and keep the setup simple to maintain.

A manual Linux installation works well on remote GPU servers because it gives direct control over Python, PyTorch, model storage, and startup behavior.

What You Need

An active SwissGPU Linux server with an NVIDIA GPU.
Working SSH access.
Python 3 and Git installed on the server.
An open port you can use for the web UI, typically 8188.
At least one compatible checkpoint model to load after installation.

Step 1: Verify GPU Availability

nvidia-smi

Confirm that the server sees the GPU before installing any Python dependencies.

Step 2: Clone ComfyUI

git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI

The official ComfyUI docs recommend cloning the repository and installing dependencies inside a dedicated environment.

Step 3: Create an Isolated Python Environment

python3 -m venv venv
source venv/bin/activate

This keeps ComfyUI dependencies separate from the rest of the system.

Step 4: Install PyTorch and ComfyUI Dependencies

On a modern NVIDIA server, install the CUDA-enabled PyTorch wheels first:

pip install torch torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/cu128

Then install ComfyUI requirements:

pip install -r requirements.txt

Step 5: Start ComfyUI

python main.py --listen 0.0.0.0 --port 8188 --enable-manager

Once it starts, openhttp://<server-host>:8188in your browser.

The --enable-manager flag turns on ComfyUI-Manager, which makes custom node installation much easier later.

Step 6: Add Models

At minimum, ComfyUI needs a checkpoint model. Common model folders are:

ComfyUI/models/checkpoints
ComfyUI/models/vae
ComfyUI/models/loras
ComfyUI/models/controlnet

The usual starting point is to place your main checkpoint file inComfyUI/models/checkpointsand refresh the browser after the file is present.

Step 7: Reuse a Shared Model Directory

If you keep models outside the ComfyUI repository, createextra_model_paths.yamlin the ComfyUI root and add entries like this:

my_models:
    base_path: /srv/ai-models
    checkpoints: checkpoints
    vae: vae
    loras: loras
    controlnet: controlnet

This is useful when you want ComfyUI to share checkpoints and LoRAs with other tools.

Step 8: Keep ComfyUI Running as a Service

If you want ComfyUI to survive shell disconnects and reboots, create a user service template:

[Unit]
Description=ComfyUI
After=network-online.target

[Service]
User=%i
WorkingDirectory=/home/%i/ComfyUI
ExecStart=/home/%i/ComfyUI/venv/bin/python main.py \
  --listen 0.0.0.0 --port 8188 --enable-manager
Restart=always

[Install]
WantedBy=multi-user.target

Save it as /etc/systemd/system/comfyui@.service, then enable it for your user:

sudo systemctl daemon-reload\nsudo systemctl enable --now comfyui@$(whoami)

Step 9: Update ComfyUI Later

cd ComfyUI
git pull
pip install -r requirements.txt

Updating regularly is useful because ComfyUI moves quickly, especially around nodes, workflows, and frontend changes.

Practical First Steps in the UI

Open the web UI and load a basic text-to-image workflow.
Select a checkpoint from the loader node.
Set a small resolution first to verify the pipeline works.
Queue one generation and monitor GPU usage with nvidia-smi.
Only after the basic workflow works should you add custom nodes or larger models.

Troubleshooting

journalctl -u comfyui@$(whoami) -f

Blank page or cannot connect: confirm the process is still running and that the selected port is reachable.
No models in the checkpoint loader: verify the files are in models/checkpoints or correctly mapped through extra_model_paths.yaml.
Torch or CUDA errors: reinstall the PyTorch wheels that match the server GPU and driver stack.
Out-of-memory errors: lower resolution, use a smaller checkpoint, or reduce batch size.

SwissGPU