Running Stable Diffusion 3.5 with ComfyUI on a Cloud GPU

Stable Diffusion 3.5 is a downloadable text-to-image model family, and I will show how I run it through ComfyUI on a cloud GPU. Vultr is sponsoring this video. While I describe the model as completely open source and free, Stability AI distributes the weights under its [Community License](https://stability.ai/news-updates/introducing-stable-diffusion-3-5), which has conditions; eligible non-commercial use is free, while commercial use can require an enterprise license depending on revenue and terms.

A cloud GPU can be an alternative if your local machine lacks sufficient capacity. I show an account promotion available at the time and deploy an Ubuntu Cloud GPU instance on Vultr, choosing one NVIDIA L40S and disabling automatic backups for the demonstration. Cloud compute is billed separately from the model license; current availability, rates, promotions, and backup requirements should be checked before provisioning.

Once the instance is running, I use SSH to connect, pull a Docker image for ComfyUI, create a storage directory, and start the container. I inspect the container logs to check that it has started. The specific image, commands, and file paths in this December 2024 walkthrough are version-dependent, so consult the current project and provider documentation.

The video opens port 8188 and accesses ComfyUI by a public IP address. That direct approach can expose an unauthenticated interface to the internet. A safer setup binds the service to the server’s loopback interface, uses an SSH tunnel or a private authenticated gateway, and restricts network access rather than opening the port broadly.

I install the Hugging Face command-line tools, sign in with a read token, and download the model files. Access can require accepting model terms. I move the downloaded model and text encoders into the directories mounted by the container, then refresh the interface. Keep tokens and server credentials private, and revoke or rotate any that have been exposed.

In ComfyUI, I enter a prompt in the text-encoding node and queue the workflow. The first image does not fully match my request, although I like how it looks. I try a second prompt, a cat with dragon wings, and show the output. The workflow and image are examples rather than proof that all prompts will be followed reliably.

I mention that users can try different workflows and models once the interface is configured. In the sponsored closing, I cite a time-limited GH200 promotion at $2.99 per hour and say it was 25% below the then-normal price. That was a historical offer, not a current price or a rate for the L40S used in the demo; check present pricing and shut down or remove unused resources to avoid ongoing charges.