How to Run ComfyUI on RunPod (2026): Official Template, Network Volume, Model Folders and Costs

If your own GPU can’t handle a workflow, say Wan 2.2 at 720p, a full BF16 image model or a long SeedVR2 upscale, renting a GPU on RunPod for an hour is often the cheapest fix. The catch is that a badly set up pod either loses your models every time it shuts down or quietly bills you all month. This is the setup I recommend in November 2026: the official ComfyUI template, a network volume so your models persist, the exact folder paths, and the habits that keep the bill small.

Why RunPod for ComfyUI?

• Per-second billing on pods, so a 40-minute session costs 40 minutes.
• Ready-made templates: the official ComfyUI image boots with Manager and common nodes already installed.
• Network volumes keep your 50-150 GB of models between sessions, so you don’t re-download every time.
• Wide GPU range, from RTX 3090s to H100 and B200.

Vast.ai is often cheaper per hour but has no equivalent to network volumes on every host, and Google Colab doesn’t allow ComfyUI’s web UI on the free tier (see Colab free GPU limits). I compare all the hosts in the cheapest cloud GPU for AI.

What It Costs

Prices vary by region, Secure vs Community Cloud and availability, so always check the price shown on the deploy screen. Example session: 2 hours of RTX 4090 plus a 100 GB volume for the month comes to roughly $2-$9 for the GPU time and $7 for storage.

Which GPU to Pick

Not sure how much memory a model needs? My VRAM requirements guide lists image and video models by size.

Step 1: Create a Network Volume

1. Sign up at runpod.io and add credit (a $10 top-up is enough to start).
2. Open Storage > New Network Volume.
3. Choose a datacenter region. Pick one that lists the GPU you want as available, because pods using this volume must run in the same region.
4. Name it, for example comfyui-models.
5. Set the size: 50 GB for a couple of image models, 100 GB for a full Wan 2.2 setup (about 80 GB of files), more for several video models. You can increase it later, but you can’t make it smaller.
6. Click Create.

Step 2: Deploy the ComfyUI Template

1. In Storage, click your volume and choose Deploy Pod (the volume is preselected).
2. Pick a GPU from the list for that region.
3. Click Change Template and search for ComfyUI. Choose the official one maintained by RunPod. For RTX 5090, RTX PRO 6000 or B200 pick ComfyUI Blackwell edition instead.
4. GPU count: 1. Pricing: On-Demand (Spot/interruptible is cheaper but can be stopped at any time).
5. Click Deploy On-Demand.
6. Open the pod’s Logs. The first boot installs ComfyUI onto your volume and can take 10-30 minutes. Later boots are much faster because everything is already on the volume. The boot log is at /workspace/comfyui.log.

Step 3: Open ComfyUI and the Terminal

• Click Connect on the pod and open HTTP Service :8188. That’s ComfyUI.
• For a terminal, use Connect > Web Terminal, or the Jupyter/VS Code service if your template exposes one (port 8888 on many templates).

Treat the 8188 URL like a password. Anyone with it can use your ComfyUI while the pod runs.

Step 4: Download Models Into the Right Folders

On the official template, ComfyUI lives at /workspace/runpod-slim/ComfyUI. Models go under:

Fastest way to download from Hugging Face in the terminal:

cd /workspace/runpod-slim/ComfyUI/models/diffusion_models
wget -c "https://huggingface.co/ORG/REPO/resolve/main/FILE.safetensors"

For gated or many files, use the Hugging Face CLI:

pip install -U huggingface_hub
hf auth login
hf download ORG/REPO FILE.safetensors --local-dir /workspace/runpod-slim/ComfyUI/models/vae

Civitai models can be pulled through the preinstalled Civicomfy node from inside ComfyUI. After downloading, press R in ComfyUI to refresh model lists. If a workflow complains about a missing file, my ComfyUI model folders cheat sheet shows where each type belongs.

Step 5: Install Custom Nodes

1. In ComfyUI open Manager > Install Missing Custom Nodes after loading a workflow, or Custom Nodes Manager to search by name.
2. Click Install, then Restart when prompted and refresh the browser.
3. Manual install from the terminal:

cd /workspace/runpod-slim/ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF
cd ComfyUI-GGUF && pip install -r requirements.txt

Then restart ComfyUI from Manager. Because custom nodes are on the volume, they’re still there next time. My list of must-have ComfyUI custom nodes is a good starting set.

To update ComfyUI itself: Manager > Update ComfyUI, or in the terminal cd /workspace/runpod-slim/ComfyUI && git pull && pip install -r requirements.txt.

Step 6: Get Your Images and Videos Out

• Outputs are saved in /workspace/runpod-slim/ComfyUI/output/ (videos in output/video on the official template).
• Download single files from the ComfyUI queue or image viewer, or from Jupyter/VS Code’s file browser.
• For a whole folder, zip it in the terminal: cd /workspace/runpod-slim/ComfyUI && zip -r outputs.zip output, then download the zip.
• Save your workflow JSON locally too (Workflow > Export), see how to save and export ComfyUI workflows.

Step 7: Stop Paying

• Stop the pod to pause GPU billing. With a network volume you can also simply Terminate it: everything under /workspace stays on the volume.
• Next time, deploy a new pod on the same volume. The pod boots in minutes with your models and nodes in place.
• Delete the network volume when the project is over, or it keeps billing every month.
• Set a low-balance alert in Billing so a forgotten pod can’t drain your card.

Troubleshooting

No GPUs available in my volume’s region

Pods must run in the volume’s region. Wait, pick another GPU type in that region, or create a new volume in a region with stock.

Port 8188 shows “not ready” or a 502 error

ComfyUI is still installing or restarting. Check Logs or tail -f /workspace/comfyui.log and wait until it prints the listening address.

“CUDA error: no kernel image is available” on RTX 5090 or B200

You used the normal template on a Blackwell GPU. Redeploy with ComfyUI Blackwell edition, or install a PyTorch build for CUDA 12.8+.

Out of disk space

Increase the network volume in Storage (it can only grow), or delete unused models. du -sh /workspace/runpod-slim/ComfyUI/models/* shows what’s taking space.

Out of memory during video generation

Use FP8 or GGUF versions of the model, lower resolution or frame count, or redeploy on a larger GPU. The volume moves with you.

Custom node import failed after restart

Open the log, find the missing Python package and run pip install package-name in the terminal, then restart ComfyUI.

Frequently Asked Questions

How much does it cost to run ComfyUI on RunPod?

Roughly $0.50-$0.90 per hour for an RTX 3090 or 4090, about $1.19 for an RTX 5090, plus $0.07 per GB per month for a network volume. A 100 GB volume is about $7 a month.

Which RunPod template should I use for ComfyUI?

The official ComfyUI template maintained by RunPod. Use the ComfyUI Blackwell edition for RTX 5090, RTX PRO 6000 and B200 GPUs.

Where do I put models on RunPod ComfyUI?

On the official template, under /workspace/runpod-slim/ComfyUI/models/, in subfolders such as diffusion_models, text_encoders, vae and loras.

Do I lose my models when I stop a RunPod pod?

Not if they are on a network volume mounted at /workspace. Files on the container disk outside /workspace are lost when the pod is terminated.

Is a network volume worth it?

Yes if you use ComfyUI more than once or twice. Re-downloading 50-100 GB of models every session wastes more GPU time than the volume costs.

What GPU is best for Wan 2.2 on RunPod?

An RTX 5090 or L40S for FP8 at 480p-720p, and an RTX PRO 6000 or H100 for 720p FP16 without offloading.

Can other people access my RunPod ComfyUI?

Anyone with the pod’s proxy URL can open it while it runs, so don’t share the link and stop the pod when you finish.

Final Thoughts

The one habit that matters most on RunPod is putting everything on a network volume and terminating pods freely. Once that’s set up, renting a 32-96 GB GPU for a single heavy job costs about the same as a coffee, and you can keep your local card for everyday work. If you’re weighing renting against buying, read RTX 5090 vs RTX 5080 for AI first.

Ready-to-run workflows to try on a rented GPU: Wan 2.2 in ComfyUI and MiniMax H3 to 1080p 60 FPS with SeedVR2.

Sources

• RunPod: Wan 2.2 in ComfyUI, GPU requirements and setup (official template, paths, volume pricing, 27 Oct 2026)
• Next Diffusion: How to set up ComfyUI on RunPod with a network volume
• RunPod docs: Network volumes
• Freakyjolly: Cloud GPU prices compared (Nov 2026)