RunPod for AI images: templates, GPUs and what it costs

RunPod rents you a GPU by the hour, so you can run big image models in your browser instead of buying an expensive graphics card. On secure cloud in October 2026, an RTX 4090 costs about $0.74 an hour and an RTX 5090 about $0.99. You deploy a template, open ComfyUI on port 8188, do the work, then stop and terminate the pod.

What is RunPod, and why rent instead of buy?

A cloud service where you rent a computer with a big GPU, billed while it runs. Image models need lots of VRAM: 24 GB to make images, more to train. Renting means you use a 96 GB card for an afternoon for a few dollars, then pay nothing until next time.

Sign up at RunPod and add a few dollars of credit.

Which RunPod GPU should you pick?

Pick by VRAM first, then price. 24 GB runs image generation and bots. 32 to 48 GB handles datasets. LoRA training in the course runs on the 96 GB RTX PRO 6000: 32 to 48 GB cards need quantization and Low VRAM, which gives a worse LoRA. When your first choice is unavailable, take another card in the same tier.

Community cloud is cheaper: an RTX 4090 was about $0.34 an hour and an RTX 5090 about $0.69 on 2 October 2026. Prices change; check RunPod’s pricing page. To add up a whole build, use the free RunPod cost calculator.

How do you run ComfyUI on RunPod?

Deploy a template, wait for it to boot, open it in your browser. The first boot is the slow part: the template downloads its models, often tens of gigabytes, and ComfyUI only opens when that’s done. After that, you load a workflow and press Run.

1. Deploy the template and pick a GPU from the table above.
2. Set passwords. If the template has JUPYTER_PASSWORD (and AI_TOOLKIT_AUTH for training), set them. Without them, anyone with your pod’s URL can open it.
3. Wait for first boot. Watch the logs until the downloads finish.
4. Connect to the HTTP port: 8188 for ComfyUI, 8888 for JupyterLab, 8675 for AI Toolkit.
5. Load the workflow .json and press Run. If a model shows as missing, put the file in the right folder.
6. Download your results from JupyterLab.
7. Stop, then terminate the pod.

What’s the difference between stop and terminate?

Stopping releases the GPU but keeps your volume disk, and you’re still charged for that disk while stopped. Terminating deletes everything that isn’t on a network volume and ends the bill. RunPod’s pod docs say exactly this. Download what you need, then terminate.

Storage prices from RunPod’s pricing docs, October 2026:

So a 100 GB volume disk left on a stopped pod costs about $20 a month for doing nothing. That’s the most common surprise bill.

Pods or Serverless: which do you need?

Pods for working: ComfyUI, datasets, training. Serverless for anything that answers requests, like a bot. A serverless endpoint is billed per second while a worker runs and scales to zero when idle, per RunPod’s serverless pricing. That’s how our free Discord face generator and the private Telegram bot in video 4 work.

The first request after a quiet spell is slower while a worker starts. FlashBoot, on by default, shortens that.

Do you need a network volume?

Not at the start. A network volume keeps files between pods at $0.07 per GB per month, but it locks you to one data center. When we tested that setup, most GPUs in the chosen data center were unavailable. For one character, downloading your LoRA and datasets to your own computer is simpler.

What can go wrong?

• The bill keeps running after you’re done. You stopped the pod but didn’t terminate it. Volume disks bill double while stopped.
• ComfyUI won’t open. First boot is still downloading models. Wait for the logs. If it never gets there, work through why a pod gets stuck on initializing.
• “CUDA driver too old” in the logs. Deploy again and set the CUDA version filter to a newer version (our trainer needs 13.0).
• Your GPU isn’t there after a restart. A stopped pod can come back without a free GPU on that machine. Deploy a new pod instead.
• Your balance hits $0. RunPod terminates pods without a network volume and the data is lost. Keep a small balance while you work.
• Someone else opens your pod. Set the template passwords before you deploy.