Rent a GPU for ComfyUI with crypto and set it up on a cloud machine
ComfyUI runs the same on a rented GPU, paid for in crypto only on Petaflare, as on your own machine: you install it once on the disk and open it in your browser. These are the commands, from renting the card to stopping the bill.
To run ComfyUI on a rented GPU, start a machine with an RTX 4090 (24 GB, $0.34 an hour) or an RTX 5090 (32 GB, $0.49 an hour) on our PyTorch image, connect over SSH, clone ComfyUI and install its requirements, then start it and open it through an SSH tunnel at localhost:8188. Models go in the ComfyUI/models folders. When you are done, stop the machine: billing stops and the disk, with your models and outputs, is kept for next time.
Which card for which ComfyUI workflow
| Workflow | Card | Per hour |
|---|---|---|
| SD 1.5 and SDXL, with LoRAs and ControlNets | RTX 4090, 24 GB | $0.34 |
| Flux.1 dev in FP8, SD 3.5 Medium | RTX 4090, 24 GB | $0.34 |
| Flux.1 dev in BF16, SD 3.5 Large, larger batches | RTX 5090, 32 GB | $0.49 |
| Wan 2.1 14B video, graphs that keep several models loaded | L40S, 48 GB | $0.87 |
Memory per model, with sources, is in Stable Diffusion VRAM needs. Whole cards only, billed by the minute.
Set up ComfyUI on a rented GPU
Rent the machine
Buy GPU-hours for the card you need, then launch one GPU with the PyTorch image and paste your SSH public key. See what happens after you pay.
Connect
When the dashboard shows the machine's address, connect with
ssh root@<address>. Runnvidia-smito see the card and the highest CUDA version its driver supports.Install ComfyUI
Clone the official repository, create a Python virtual environment and install PyTorch and ComfyUI's requirements (commands below).
Start it inside tmux
tmuxkeeps ComfyUI running when your SSH connection drops or your laptop sleeps.Open the tunnel
From your own computer, forward port 8188 over SSH and open
http://localhost:8188. Nothing is exposed to the internet.Add models
Download checkpoints, diffusion models, text encoders, VAEs and LoRAs into their folders under
ComfyUI/models.Stop the machine
Stop it from the dashboard when you finish. It stops using GPU-hours, and its disk keeps your setup for next time.
# On the rented machine: install ComfyUI
apt-get update && apt-get install -y git tmux python3-venv
git clone https://github.com/Comfy-Org/ComfyUI.git
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
pip install -r manager_requirements.txt # for ComfyUI-Manager# Still on the machine: start ComfyUI in a tmux session
tmux new -s comfy
source ~/ComfyUI/venv/bin/activate
cd ~/ComfyUI && python main.py --enable-manager
# Detach with Ctrl+B then D; reattach later with: tmux attach -t comfy
# On your own computer: open a private tunnel, then browse to http://localhost:8188
ssh -N -L 8188:127.0.0.1:8188 root@<address>Where models go
| Folder under ComfyUI/models | What goes there | Example |
|---|---|---|
| checkpoints | Full checkpoints with model, encoder and VAE in one file | SD 1.5, SDXL |
| diffusion_models | The image or video model alone | Flux.1 dev, Wan 2.1 |
| text_encoders | CLIP and T5 text encoders | clip_l, t5xxl in FP8 or FP16 |
| vae | Image decoders | The Flux or SDXL VAE |
| loras | LoRA adapters | Style or character LoRAs |
Each model's page on Hugging Face says which files it needs. ComfyUI rescans the folders when you refresh the interface.
cd ~/ComfyUI/models
# An open checkpoint: download it directly
wget -P checkpoints https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/resolve/main/sd_xl_base_1.0.safetensors
# A gated model such as Flux.1 dev: accept its licence on Hugging Face first, then
pip install -U huggingface_hub
hf auth login
hf download black-forest-labs/FLUX.1-dev flux1-dev.safetensors --local-dir diffusion_modelsPaying in crypto, only while you generate
A rented machine is billed for every minute it runs, whether ComfyUI is busy or waiting for your next prompt. There is no per-image billing. So the habit that saves money is simple: stop the machine when you step away, and start it again when you come back. The disk keeps ComfyUI, your custom nodes and your models, so a restart takes you straight back to work. Prepaid GPU-hours never expire, which makes short sessions spread over weeks cheap: ten one-hour sessions on an RTX 4090 cost $3.40. Prices for every card are on cloud GPU pricing. The hours themselves are bought in blocks paid in crypto, from $25.50 for 75 hours of RTX 4090.
Video generation in ComfyUI
Video models are where a rented GPU makes the most difference, because they need more memory than most people own. Wan 2.1 comes in two sizes: the 1.3B text-to-video model needs 8.19 GB of VRAM according to its authors, and runs on any card we rent; the 14B model needs about 28 GB for its weights in BF16 or 14 GB in FP8, plus its encoders. For it, rent an RTX 5090 and use FP8 weights, or take an L40S with 48 GB to keep everything in memory. A clip often takes minutes rather than seconds, so think in cost per clip: at 4 minutes a clip, an L40S costs about 6 cents per clip.
Keep your work
Stopping a machine keeps its disk; deleting it wipes everything. Before you delete a machine, copy your outputs and save your workflows from the ComfyUI menu as JSON files.
# On your own computer: copy all outputs from the machine
rsync -avP root@<address>:~/ComfyUI/output/ ./comfyui-output/ComfyUI cloud GPU questions
What GPU is needed for ComfyUI?
Any NVIDIA card with 8 GB runs SD 1.5 workflows. For SDXL with LoRAs and Flux in FP8, 24 GB is the comfortable minimum; for Flux in BF16 and SD 3.5 Large, 32 GB; for 14B video models, 48 GB or FP8 weights.
How much does it cost to rent a GPU for ComfyUI?
$0.34 an hour for an RTX 4090, $0.49 for an RTX 5090 and $0.87 for an L40S, billed by the minute from prepaid blocks. The smallest block is 75 hours on an RTX 4090 for $25.50.
Can I install custom nodes?
Yes. You have root on the machine, so ComfyUI-Manager or a git clone into ComfyUI/custom_nodes works as it does at home.
Is ComfyUI exposed to the internet?
Not with this setup. ComfyUI listens only on the machine itself, and you reach it through an SSH tunnel that only your key can open.
Sources
- ComfyUI, official repository and READMEread 19 November 2026
- ComfyUI-Manager repositoryread 19 November 2026
- Hugging Face CLI documentationread 19 November 2026
- Wan 2.1 repositoryread 19 November 2026