If your laptop has no capable GPU, rent one by the hour: deploy Runpod's official ComfyUI template on a 24 GB card such as an RTX 4090, open port 8188 in your browser, and put your models on a network volume so you do not download them again next time. A 4090 pod listed at $0.89 an hour on Runpod's pricing page on 2026-10-16, and a 100 GB network volume costs about $7 a month to keep.

The setup takes a few clicks. What costs money is storage and a pod left running, so this guide spends most of its time there.

> **Key Takeaways**
>
> - **A 24 GB card covers most SDXL and Flux work.** Runpod lists the RTX 4090 and RTX 3090 with 24 GB of VRAM at $0.89 and $0.50 an hour on its Secure Cloud ([Runpod pricing](https://www.runpod.io/pricing), retrieved 2026-10-16).
> - **There is an official template.** Runpod's "ComfyUI" template runs the `runpod/comfyui:latest` image and exposes the UI on port 8188 ([Runpod docs, ComfyUI tutorial](https://docs.runpod.io/tutorials/pods/comfyui), retrieved 2026-10-16).
> - **Stopping is not free.** A stopped pod's volume disk is billed at $0.20/GB/month, double the running rate ([Runpod docs, pod pricing](https://docs.runpod.io/pods/pricing), retrieved 2026-10-16).
> - **Network volumes are the cheaper way to keep models.** They cost $0.07/GB/month under 1 TB and survive termination ([Runpod pricing](https://www.runpod.io/pricing), retrieved 2026-10-16).
> - **Flux in fp8 halves memory.** ComfyUI's own example says the fp8 weight type "will lower the memory usage by half but might reduce quality a tiny bit" ([ComfyUI examples, Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/), retrieved 2026-10-16).

## Why cloud GPU for ComfyUI

ComfyUI can run without a GPU. Its system requirements page documents a `--cpu` mode ([ComfyUI docs, system requirements](https://docs.comfy.org/installation/system_requirements), retrieved 2026-10-16). For SDXL or Flux that is a way to check a workflow loads, not a way to make images.

The models are the problem. Flux.1 [dev] is a "12 billion parameter rectified flow transformer" stored in BF16 ([Hugging Face, FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev), retrieved 2026-10-16). At two bytes per parameter, the transformer weights alone are about 24 GB before the text encoders and VAE load. That is our arithmetic, not a vendor figure, but it explains why an 8 GB laptop GPU struggles and why a rented 24 to 48 GB card does not.

Renting makes sense when:

- You generate in bursts (an evening a week, a client project) rather than all day.
- You want to try Flux or video workflows before buying a high-end card.
- You need more VRAM than any consumer card has, for a short job.

If you run ComfyUI all day, every day, the hourly bill starts to look like a hardware payment.

### Your options in one table

| Option | How you pay | Price read on 2026-10-16 | Best for | Not for |
|---|---|---|---|---|
| [Runpod](aff:runpod) | Per hour, plus storage | RTX 4090 $0.89/hr Secure Cloud; network volume $0.07/GB/mo | People who want full control and the cheapest dedicated GPU with a ready template | Anyone who never wants to think about storage or shutdown |
| [Vast.ai](https://vast.ai/pricing) | Marketplace, per second | "Prices are set by the market, not by Vast" | Price hunters comfortable picking hosts | Beginners; interruptible instances can be reclaimed |
| [RunComfy](https://www.runcomfy.com/pricing) | Per second, optional Pro plan | Large (24 GB) $1.75/hr pay as you go, $1.39/hr on Pro ($19.99/mo) | People who want ComfyUI managed for them | Heavy users who will pay the managed-service markup every hour |
| [Google Colab](https://colab.research.google.com/) | Compute units | Free tier restricts web UIs | Notebook users | Running the ComfyUI web UI on the free tier |

Vast.ai's pricing page offers on-demand, interruptible ("50%+ cheaper") and reserved instances, with "no minimum hours or rounding up" ([Vast.ai pricing](https://vast.ai/pricing), retrieved 2026-10-16). Its live GPU prices did not load when we read the page, so they are not quoted here.

Colab's FAQ says free runtimes do not allow "bypassing the notebook UI to interact primarily via a web UI", which is how ComfyUI works, and that paid plans or a positive compute unit balance remove that restriction ([Google Colab FAQ](https://research.google.com/colaboratory/faq.html), retrieved 2026-10-16). We could not open Google's consumer price page for Colab Pro, so its prices are left out.

## Choosing a GPU (VRAM guide)

Pick by VRAM first, then by price. None of the official model cards or ComfyUI docs we read publishes a minimum VRAM figure for SDXL or Flux, so the tiers below are sizing guidance from parameter counts and ComfyUI's fp8 note, not vendor requirements.

| GPU on Runpod | VRAM | Secure Cloud $/hr | Reasonable fit |
|---|---|---|---|
| RTX 3090 | 24 GB | $0.50 | SDXL, Flux in fp8, learning ComfyUI cheaply |
| RTX 4090 | 24 GB | $0.89 | Same as 3090, faster; the default pick |
| RTX 5090 | 32 GB | $1.19 | Flux with more headroom; needs the Blackwell template |
| A40 / L40S | 48 GB | $0.59 / $1.09 | Flux in full precision, large batches, LoRA training |
| A100 / H100 | 80 GB | $1.79 / from $2.89 | Video models and jobs that will not fit elsewhere |

Prices and VRAM from [Runpod pricing](https://www.runpod.io/pricing), retrieved 2026-10-16. The same page's structured data also lists lower Community Cloud rates (for example $0.34/hr for an RTX 4090 and $0.22/hr for an RTX 3090). Community Cloud runs on third-party hosts; check the rate shown in the deploy screen before you start, because it moves with supply.

The A40 is cheaper than the 4090 per hour on Secure Cloud and has twice the VRAM, so if a Flux workflow runs out of memory on 24 GB, try it before jumping to an A100. And system RAM matters for Flux too: ComfyUI recommends the fp16 T5 text encoder only "if you have more than 32GB ram" ([ComfyUI examples, Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/), retrieved 2026-10-16). Runpod lists 41 GB of system RAM with its 4090 and 125 GB with its 3090.

## Deploy the ComfyUI template on Runpod

These steps follow Runpod's ComfyUI tutorial ([Runpod docs](https://docs.runpod.io/tutorials/pods/comfyui), retrieved 2026-10-16). Plan on 15 minutes for the clicks, though Runpod warns the first start can take longer.

1. **Create a network volume first** (Storage, then New Network Volume). Pick a data center that has the GPU you want, and size it for your models: the Flux dev transformer alone is roughly 24 GB at BF16, so 100 GB leaves room for SDXL, Flux and a few LoRAs. A network volume lives in one data center, so creating it first keeps the pod and the volume together.
2. **Deploy a pod.** Choose your GPU, then pick the **ComfyUI** template. For an RTX 5090 or B200, pick **ComfyUI Blackwell Edition** instead, which uses the `runpod/comfyui:cuda12.8` image.
3. **Attach the network volume** and choose **On-Demand** pricing.
4. **Wait for it to start.** Runpod's tutorial says setup can take up to 30 minutes, and if you see "Not Ready" or "Bad Gateway", wait two to three minutes and refresh.
5. **Connect.** Click Connect, then "Connect to HTTP Service [Port 8188]". ComfyUI opens in a browser tab. The template also exposes port 8080 for a file browser and 22 for SSH.

## Install models and custom nodes

### Models

ComfyUI looks for models in fixed folders under `models/`: `checkpoints`, `loras`, `vae`, `controlnet`, `clip`, `clip_vision`, `diffusion_models`, `embeddings` and `upscale_models` ([ComfyUI docs, manual install](https://docs.comfy.org/installation/manual_install), retrieved 2026-10-16). On the Runpod template, Runpod's docs give the path as `/workspace/madapps/ComfyUI/models`.

For Flux dev, ComfyUI's example puts the files here ([ComfyUI examples, Flux](https://comfyanonymous.github.io/ComfyUI_examples/flux/), retrieved 2026-10-16):

- `flux1-dev.safetensors` in `models/diffusion_models/`
- `t5xxl_fp16` or `t5xxl_fp8_e4m3fn_scaled`, plus `clip_l`, in `models/text_encoders/`
- `ae.safetensors` in `models/vae/`
- Or the single-file `flux1-dev-fp8.safetensors` in `models/checkpoints/`, the simplest option on a 24 GB card

Flux.1 [dev] is gated on Hugging Face: you accept the license before you can download it, and it is released under a non-commercial license ([Hugging Face, FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev), retrieved 2026-10-16). Read that license before you use outputs for client work. To download on the pod, open a terminal (SSH or the template's web terminal) and use `wget` or the Hugging Face CLI with an access token, straight into the folders above.

### Custom nodes

ComfyUI Manager is now part of ComfyUI core and handles node installs and model downloads ([ComfyUI docs, Manager](https://docs.comfy.org/manager/install), retrieved 2026-10-16). Runpod's tutorial also points to Manager's Model Manager as the in-browser way to fetch models. When you load a shared workflow with red "missing" nodes, open Manager, install the missing nodes, and restart ComfyUI when Manager prompts you.

Install custom nodes into the network volume path, not the container disk, or you will reinstall them every session.

## Persistent storage and costs

Runpod has three kinds of storage, and choosing the wrong one is the most common surprise on the bill ([Runpod docs, pod pricing](https://docs.runpod.io/pods/pricing), retrieved 2026-10-16):

| Storage | Running | Stopped | What survives |
|---|---|---|---|
| Container disk | $0.10/GB/mo | Not charged | Nothing; "erased when Pod stops" |
| Volume disk | $0.10/GB/mo | $0.20/GB/mo | Kept until the pod is deleted |
| Network volume | $0.07/GB/mo | $0.07/GB/mo | Kept after the pod is terminated; can attach to new pods |

The network volume is the right home for models. It costs less than a stopped volume disk, and you can terminate the pod completely between sessions instead of leaving it parked.

### What a month costs

A worked example using the prices above, for someone generating 10 hours a week on an RTX 4090 Secure Cloud pod with a 100 GB network volume:

- GPU: 40 hours × $0.89 = $35.60
- Network volume: 100 GB × $0.07 = $7.00
- Total: about $42.60 a month

Run the same hours on a 3090 and the GPU line drops to $20.00. Leave a 100 GB volume disk on a stopped pod instead of using a network volume, and that line becomes $20.00 a month for storage you are not using.

Runpod's docs add a detail worth knowing: if your balance reaches $0, pods with a network volume are stopped and keep their data, while pods without one are terminated and their data cannot be recovered. Savings plans cover GPU compute only, not storage.

## Shut down to avoid bills

The GPU meter runs while the pod runs, whether ComfyUI is generating or idle in a forgotten tab. Runpod's tutorial says it plainly: "Stop your Pod when you're finished to avoid unnecessary charges."

Know the difference between the two buttons ([Runpod docs, manage pods](https://docs.runpod.io/pods/manage-pods), retrieved 2026-10-16):

- **Stop** "releases the GPU and preserves data in `/workspace`", but "you'll still be charged for volume disk storage while stopped". When you restart, "you may be allocated zero GPUs if capacity has changed".
- **Terminate** "permanently deletes all data not stored in a network volume".

With models on a network volume, the clean routine is: save your workflow JSON and download your outputs, then terminate the pod. Next session, deploy a fresh pod on the same volume. That also avoids the zero-GPU problem, because you pick whatever card is free.

A checklist before you close the tab:

1. Outputs downloaded or saved to the network volume.
2. Workflow exported (Workflow menu, Export).
3. Pod terminated, or stopped if you use volume disk.

If you want ideas for what to build once the pod is running, our [FLUX.1 Kontext review](/blog/flux1-kontext-ai-editor) covers one of the editing models that runs in ComfyUI, and [the creative edge of an AI image generator](/blog/ai-image-generator-creativity) looks at where generated images fit in real work.

## Frequently Asked Questions

### How much does it cost to run ComfyUI in the cloud?

On Runpod, an RTX 4090 pod listed at $0.89 an hour on Secure Cloud and an RTX 3090 at $0.50 an hour on 2026-10-16. Add storage: a network volume is $0.07 per GB per month. Ten hours a week on a 4090 with 100 GB of storage comes to roughly $43 a month.

### What GPU do I need for Flux in ComfyUI?

A 24 GB card such as an RTX 3090 or 4090 is the practical starting point when you use the fp8 Flux checkpoint, which ComfyUI says halves memory use. For full-precision Flux, larger batches or LoRA training, a 48 GB A40 or L40S gives headroom. No official minimum is published, so treat these as sizing guidance.

### Can I run ComfyUI on Google Colab for free?

Not comfortably. Colab's FAQ lists "bypassing the notebook UI to interact primarily via a web UI" as disallowed on free runtimes, and ComfyUI is a web UI. Google says paid plans or a positive compute unit balance remove that restriction.

### Will I lose my models when I stop the pod?

Not if they are on a volume disk or network volume. Container disk is erased when a pod stops. Volume disk survives a stop but is deleted when you terminate. A network volume survives termination and can be attached to a new pod, which is why it is the best place for models.

### Is RunComfy cheaper than Runpod?

Per hour, no. RunComfy's 24 GB "Large" machine listed at $1.75 an hour pay as you go, or $1.39 with the $19.99 Pro plan. Runpod's 24 GB cards listed from $0.50 to $0.89 on Secure Cloud. RunComfy manages the setup for you, which is what the difference pays for.

## How this guide was researched

We read Runpod's pricing page and its docs on pod pricing, pod management and the ComfyUI tutorial; ComfyUI's official docs and Flux examples; the FLUX.1-dev model card on Hugging Face; and the pricing pages or FAQs for Vast.ai, RunComfy and Google Colab, all on 2026-10-16. Tools were compared from documentation, pricing and public information, not hands-on testing, and the cost example is arithmetic on listed prices. LogicBalls may earn a commission from some links in this guide, which does not change the order or the recommendations.