Running ComfyUI in the cloud: a managed service or your own GPU

You can run ComfyUI online in three ways. Use a managed service such as Comfy Cloud, a GPU cloud with ComfyUI ready to start, or a GPU server you set up.

A managed service charges only while a workflow runs. A rented GPU charges for every hour it is on, including the time you spend adjusting prompts. Which costs less depends on how long the card would sit idle.

The facts below come from Comfy Cloud's pricing page, ComfyUI's own documentation and the model makers' memory figures. We read them all on 27 October 2026.

Three ways to run ComfyUI online

The middle route is a GPU rented by the hour from a cloud that offers a ready-made ComfyUI image. It saves the set-up, but the meter runs as on any rented GPU.

Comfy Cloud: what you get

Comfy Cloud is the hosted service of Comfy Org, the company behind ComfyUI. Its pricing page says it runs on Blackwell RTX PRO 6000 cards with 96 GB of memory.

• You are charged only for active GPU time while a workflow runs; time spent building workflows uses none, by the pricing page.
• Each plan includes a monthly pool of credits. Credits you buy on top roll over for up to a year.
• One run may last up to 30 minutes, or 1 hour on the Pro plan; longer jobs are cancelled.
• More than 900 models come pre-installed. Importing your own models and LoRAs from CivitAI or Hugging Face needs the Creator, Pro or Team plan.
• New users get five free runs on real GPUs.

Its plans and prices are on its own pricing page, linked in the sources below.

Paying while you think: idle time

With ComfyUI, much of a session is spent changing prompts and settings between runs. On Comfy Cloud that time is free; on a rented GPU the meter keeps running.

So compare two numbers. The GPU minutes your workflows actually run decide the managed cost. The hours the server is switched on decide the rented cost.

A rented card wins when it would be busy most of the time it is on, or when you need runs longer than 30 minutes. With hourly billing, end each session by saving your work and deleting the server.

How much GPU memory ComfyUI needs

The card's memory decides which models fit. By the makers' own figures, collected in our GPU memory guide for image and video models:

• SDXL works with 8 GB, says Stability AI.
• FLUX.2 [klein] 4B needs about 13 GB, says Black Forest Labs.
• Wan 2.2 TI2V-5B peaked at 22.9 GB for a 720p clip on an RTX 4090, in Wan's own table.
• Wan 2.2 T2V-A14B peaked at 59.8 GB for a 720p clip on an H100, so it needs an 80 GB card.

Comfy Cloud's 96 GB cards hold all of these. On a rented GPU, 24 GB covers most image work and the smaller video models; large video models need 48 to 80 GB.

Set up ComfyUI on a rented GPU step by step

These steps follow ComfyUI's README for Linux with an NVIDIA card. They assume Ubuntu with the NVIDIA driver installed.

1. Install ComfyUI

ComfyUI's README asks for a PyTorch build for CUDA 13.0 or newer on RTX 20 series cards and later:

sudo apt install -y git python3-venv
git clone https://github.com/Comfy-Org/ComfyUI
cd ComfyUI
python3 -m venv venv && source venv/bin/activate
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130
pip install -r requirements.txt

ComfyUI's README also offers comfy-cli, which does the same with pip install comfy-cli and comfy install.

2. Add a model

Checkpoints go in the models/checkpoints folder. This downloads SDXL base 1.0 from Stability AI, a 6.9 GB file:

wget -P models/checkpoints \
  https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors

3. Start it and open it safely

By default ComfyUI listens only on the server itself, on port 8188, according to its settings file. Start it, then reach it from your own computer through an SSH tunnel, without opening a port:

python main.py
# on your own computer:
ssh -L 8188:127.0.0.1:8188 user@your-server    # then open http://localhost:8188

Run it inside tmux or screen if it should keep going after you close the SSH window.

4. Add custom nodes with ComfyUI-Manager

ComfyUI-Manager installs and updates custom nodes. The README turns it on with two steps:

pip install -r manager_requirements.txt
python main.py --enable-manager

Which route for which job

• A few images now and then, with no set-up: Comfy Cloud, which charges only while workflows run.
• Your own models and LoRAs: Comfy Cloud's Creator plan or above, or any rented GPU.
• Video runs longer than 30 minutes, or 1 hour on Pro: a rented GPU, which has no time limit per run.
• Daily heavy use, or images you cannot send to a service: a GPU server by the month.

To teach Flux a face, product or style, see our guide to training a Flux LoRA on a rented GPU. Our GPU memory guide for image and video models has the full tables.

How we checked

We read Comfy Cloud's pricing page, ComfyUI's README, its settings file and its latest release, and the model makers' pages on 27 October 2026. The memory figures are the makers' own, as collected in our memory guide.

Our own measurements will be added here, with their dates, when our test bench runs them.