Run ComfyUI on a cloud GPU
Published 27 October 2026 · prices on this page are live
ComfyUI runs best on an NVIDIA GPU with plenty of memory, and the right amount depends on the model: a card that is fine for SDXL runs out of room for FLUX or video. Renting by the second lets you use a 24 GB or 80 GB card for the evening you need it and stop paying when you close the tab.
How much GPU memory ComfyUI needs
Rules of thumb for generating, with the model's own settings. Higher resolutions, batches, ControlNets and extra LoRAs all need more.
| Model | GPU memory |
|---|---|
| Stable Diffusion 1.5 | about 4–6 GB |
| SDXL, 1024×1024 | about 8–12 GB; more with a refiner or ControlNet |
| FLUX.1 [dev] or [schnell] | about 24 GB at full 16-bit precision; about 12–16 GB with fp8 or GGUF weights |
| Video models such as Wan and HunyuanVideo | 24 GB with fp8 weights at modest resolutions; 48–80 GB for higher resolutions and longer clips |
| Training LoRAs (Kohya) | about 12–16 GB for SDXL, about 24 GB for FLUX |
The cheapest GPU for each memory size today:
| Memory | Cheapest GPU today | USD | INR | NPR |
|---|---|---|---|---|
| 8 GB or more | RTX 2060 | USD 0.08 | INR 8 | NPR 12 |
| 12 GB or more | RTX 2060 | USD 0.08 | INR 8 | NPR 12 |
| 16 GB or more | RTX 2080 Ti | USD 0.13 | INR 12 | NPR 19 |
| 24 GB or more | Tesla P40 | USD 0.18 | INR 17 | NPR 26 |
| 48 GB or more | Quadro RTX 8000 | USD 0.39 | INR 37 | NPR 58 |
| 80 GB or more | GB10 | USD 0.50 | INR 47 | NPR 74 |
What generating images costs
A thousand SDXL images at 1024×1024 and 30 steps, on today's cheapest hosts:
| GPU | Typical time | USD | INR | NPR |
|---|---|---|---|---|
| RTX 3090 | 2.4 h | USD 0.72 | INR 68 | NPR 106 |
| RTX 4090 | 1.2 h | USD 0.72 | INR 68 | NPR 106 |
| A100 SXM4 | 1.2 h | USD 0.78 | INR 74 | NPR 116 |
| RTX 5090 | 1.1 h | USD 0.83 | INR 77 | NPR 123 |
| RTX A5000 | 2.8 h | USD 1.04 | INR 98 | NPR 154 |
Prices per hour for the cards most people pick for ComfyUI:
| GPU | Memory | USD | INR | NPR |
|---|---|---|---|---|
| RTX 3090 | 24 GB | USD 0.30 | INR 28 | NPR 44 |
| RTX 4090 | 24 GB | USD 0.60 | INR 56 | NPR 88 |
| RTX A6000 | 48 GB | USD 0.65 | INR 61 | NPR 96 |
| L40S | 45 GB | USD 0.86 | INR 80 | NPR 128 |
| A100 SXM4 | 80 GB | USD 0.65 | INR 61 | NPR 96 |
| H100 SXM | 80 GB | USD 3.74 | INR 349 | NPR 556 |
Start ComfyUI on PowerPod
- Create an account and add credits by card.
- Choose the ComfyUI software, then a GPU with enough memory for your model from the table above.
- Open ComfyUI from the console. It runs in your browser; no SSH key or install is needed. JupyterLab and a terminal come with it.
- Download the models you use into ComfyUI's models folders (checkpoints, loras, vae and so on) from Hugging Face or Civitai. Data-centre hosts download them fast.
- Install custom nodes as you would at home, with ComfyUI Manager or by cloning them into custom_nodes.
- Stop the machine when you are done. A stopped machine keeps its disk, models and outputs, and bills only for that storage; start it again later from the console.
Your own GPU, a hosted ComfyUI, or a rented GPU
- Your own GPU has no hourly bill once you own it, but you are limited to its memory. An RTX 4090 launched at US$1,599.
- Hosted ComfyUI services, including one from Comfy Org, the team behind ComfyUI, run the machine for you and charge in their own way; check their pricing pages.
- A rented GPU gives you the whole machine: any model, any custom node, root access, and the card you choose for each job, paid by the second. You look after your own files, and a host can go offline, so keep copies of anything you cannot lose.
Spend less
- Build and test your workflow at low resolution on a cheaper card, then rent the big one for the final batch.
- Use fp8 or GGUF versions of FLUX and video models to fit a 16 GB or 24 GB card.
- Queue a batch and let it run; set a timer when you rent, and the idle stop catches a forgotten machine.
Questions
Which GPU should I rent for ComfyUI?
For SDXL, a 12 GB card or larger. For FLUX at full precision and for video models, 24 GB, such as the RTX 3090 or RTX 4090. For higher-resolution video, 48–80 GB, such as the L40S, A100 or H100.
Can I install custom nodes and my own models?
Yes. You have the whole machine: install nodes with ComfyUI Manager or git, and download any model from Hugging Face or Civitai into ComfyUI’s models folders.
Do I pay while ComfyUI is idle?
Yes, while the machine is running. Stop it when you finish, or let a timer or the idle stop do it; a stopped machine bills only for its disk.
Do I need to know Linux?
No. ComfyUI opens in your browser from the console. A terminal and JupyterLab are there if you want them.