Make AI images on a rented GPU: Stable Diffusion, SDXL and Flux

Image models such as Stable Diffusion XL and Flux want a graphics card with plenty of memory. Renting one by the hour lets you generate at full quality without buying a card, and you stop paying when you stop. There is no one-click image app yet, but two ways work well today: a notebook with the diffusers library, or ComfyUI or Forge running on the rented GPU.

Which GPU to rent

Memory (VRAM) decides what runs:

A 24 GB card runs all of them: an RTX 4090 at about $0.49 an hour, or an RTX 3090 at about $0.31 — a little slower, and plenty for most work. For SDXL alone, a 12 GB card is enough. Filter the marketplace by memory; more numbers are in How much GPU memory do I need?

Way 1: a notebook with diffusers

Rent a card with the PyTorch 2.4 + JupyterLab environment (you pick it at checkout). When the rental starts, open JupyterLab from Web access on the rental page, create a notebook and run:

%pip install -q diffusers transformers accelerate

import torch
from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "stabilityai/stable-diffusion-xl-base-1.0",
    torch_dtype=torch.float16, variant="fp16",
).to("cuda")

image = pipe("a lighthouse on a cliff at sunset, oil painting",
             num_inference_steps=30).images[0]
image.save("/workspace/lighthouse.png")

The first run downloads the model, about 7 GB; after that, each picture takes seconds. Change the prompt and run the last two lines again. Flux works the same way with diffusers' FluxPipeline; on a 24 GB card, call pipe.enable_model_cpu_offload() instead of .to("cuda") so that it fits.

Way 2: ComfyUI or Forge

Prefer a visual tool? ComfyUI (node-based, port 8188) and Stable Diffusion WebUI Forge (port 7860) run on the rented GPU and open in your own browser.

Choose the image. Turn on Pro mode (the switch in the header). At checkout, open Your own Docker image and type a public ComfyUI or Forge image — or the plain PyTorch image, pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime, and let the start command install ComfyUI. Under Run when it starts, for example:

apt-get update && apt-get install -y git
git clone https://github.com/comfyanonymous/ComfyUI /workspace/ComfyUI
pip install -r /workspace/ComfyUI/requirements.txt
python -c "from huggingface_hub import hf_hub_download as get; get('stabilityai/stable-diffusion-xl-base-1.0', 'sd_xl_base_1.0.safetensors', local_dir='/workspace/ComfyUI/models/checkpoints')"

Press Use this image, then pay. The command runs once, first, and its output appears in the Console on the rental page. With a ready-made image instead, its own page tells you where the app lives and which command starts it.

Start it and open it. The rental starts your image with a shell rather than with the image's own start-up, so you launch the app yourself. With the gpurent command line (123 is your rental's number), connect with ComfyUI's port forwarded to your computer, then start it:

gpurent ssh 123 -L 8188:127.0.0.1:8188
cd /workspace/ComfyUI && python main.py --listen 0.0.0.0 --port 8188

Then open http://127.0.0.1:8188 in your browser. Forge works the same way on port 7860. To forward a port without SSH, use gpurent tunnel 123 8188.

Keep your pictures

Files in /workspace are deleted when the rental ends. Download your images from the rental's Files panel, or copy them to My Drive — 2 GB that stays with your account between rentals. The rental page can also save everything to My Drive automatically before the time runs out. ComfyUI puts its pictures in /workspace/ComfyUI/output.

What it costs

You pay the host's hourly price, locked when you book; stop early and the unused time comes back as credit. An evening of SDXL — three hours on an RTX 3090 — costs about 3 × $0.31.

FAQ

Do I need to know Python?

Barely. In the notebook you only change the prompt and run the cell again; ComfyUI and Forge are point-and-click once they are running.

Can I use my own checkpoints and LoRAs?

Yes. Upload them in the Files panel or with scp, or paste a link in the Files panel to have the machine download them itself. For ComfyUI, checkpoints go in models/checkpoints and LoRAs in models/loras.

Why is the first picture slow?

The model has to be downloaded to the machine first — several gigabytes. Every picture after that comes in seconds.

Are my prompts and images private?

The model runs on the host's machine inside a sandbox (a Docker container); the host cannot open your files through the app, and nothing is sent to an AI company. As with any rented computer, do not put secrets on it that you could not afford to lose.