The ComfyUI template gives me a blank node graph in the browser: launch it, open the private link, and ComfyUI is sitting there with no models and no history, ready for whatever workflow I bring to it.
What the template installs
The template runs deploygpu/deploygpu:comfyui, an image DeployGPU builds from RunPod's own ComfyUI base image, about 5.3 GB, on a GPU VM from QuantaCloud, Massed Compute, Verda or Hyperstack. ComfyUI-Manager comes with it, so installing custom nodes is a menu inside the app rather than a git clone over SSH. Nothing else is preloaded: no checkpoints, no LoRAs, no VAEs, and no login screen of its own, because the private link's DeployGPU sign-in is already the gate in front of it.
That link is private to whoever launched it: only the DeployGPU account that launched the deployment can open it, signed in, in a browser. It cannot be shared with a teammate, has no custom domain, exposes one port and does not accept API keys, so a script cannot call it directly.
Pick a GPU for the workflow, not the app
ComfyUI itself is light. What decides the GPU is the checkpoint and the resolution in the workflow you load, the same as running that model outside ComfyUI would. Two families anchor most workflows people bring to a node graph: still-image diffusion models such as FLUX.1-dev, and newer video models such as Wan2.2.
| Model | Memory to run at full precision | GPU that fits it |
|---|---|---|
| FLUX.1-dev (text encoder, transformer and VAE together) | about 31.4 GB in BF16, by Hugging Face's own breakdown | A 48 GB card such as the L40S or RTX A6000, with room left for high-resolution work |
| FLUX.2, loaded with no CPU offload | more than 80 GB, per Hugging Face's FLUX.2 announcement | An 80 GB or larger GPU |
| Wan2.2 TI2V-5B (video) | at least 24 GB, per the model's own repository | A 48 GB card runs it with room to spare |
| Wan2.2 A14B (video) | at least 80 GB, per the model's own repository | An 80 GB A100 or H100, or the RTX PRO 6000 Blackwell |
The FLUX.1-dev figure is from the Hugging Face diffusers-quantization blog, and both Wan2.2 figures are from the Wan2.2 repository, all checked on 28 September 2026. Quantized checkpoints, CPU offload and VAE tiling all lower the real number, and ComfyUI exposes controls for most of them, so treat the table as a ceiling, not a requirement. My default for a still-image workflow is a 48 GB card, and I only move up to 80 GB or more once a workflow's own documentation says the smaller card will not hold it.
Opens the deploy page on the cheapest 1x L40S offer at 2026-10-15 18:46 UTC: Massed Compute, $0.97/hr. You can pick another GPU there before paying.
Launch it
- Click the button above, or open the marketplace, pick any VM offer and click Launch. You can also start from Templates in the dashboard sidebar: choose ComfyUI, then one of the suggested GPUs.
- On the deploy page, choose ComfyUI under Template if it is not already selected.
- Check the cost summary and the SSH key, then deploy. The first hour is reserved from your credit balance.
- Wait on the Deployments page until the status reads Running, typically 6 to 8 minutes for an app template.
- Click Open Application.
A failed launch refunds the reserved hour automatically, so trying another offer costs nothing extra.
Adding models
ComfyUI-Manager is the easy way in: open its menu inside the app, search for a checkpoint or node pack, and install it without touching SSH. For a model ComfyUI-Manager does not carry, or a large file I already have a direct link for, I go in over SSH and download straight into the container, because the container's own disk is the only disk here.
The image's exact folder layout is not something I have confirmed line by line, so I look it up rather than guess a path from another RunPod-based image:
CID=$(sudo docker ps -q --filter ancestor=deploygpu/deploygpu:comfyui)
sudo docker exec -it "$CID" ls /
Follow the directories down from there. ComfyUI keeps checkpoints, LoRAs and VAEs under a models folder with one subfolder per type, and once ls shows that folder, downloading straight into it or running docker cp from outside the container both work the same as on any other Linux machine.
Using it from outside the browser
The private link needs a signed-in browser, so a script cannot open it the way it opens for me. When a workflow needs the ComfyUI API instead of the canvas, the same SSH tunnel pattern that works for any DeployGPU app applies: forward the container's port over SSH and talk to localhost on your own machine.
ssh -N -L 8188:127.0.0.1:8188 ubuntu@YOUR_VM_IP -p 22
ComfyUI listens on port 8188 inside the container. Use the user and port from the SSH command shown on your deployment: ubuntu on QuantaCloud and Hyperstack, Ubuntu with a capital U on Massed Compute, and root on Verda.
What it costs
You pay the hourly rate of the GPU you launch on, shown in the button caption above and in the live 48 GB GPU table, and nothing extra for the template. A session that runs 2 hours and 15 minutes bills two full hours plus the started third, then refunds the unused part of that third hour when you stop. Credits are bought by card, with a 5% + $0.30 processing fee per top-up. The pricing page has the rest of the mechanics.
There is no idle timer, so a workflow left queued overnight bills all night whether or not it is still generating anything.
When something goes wrong
| What you see | Likely cause | Fix |
|---|---|---|
| The model list is empty in a workflow | Nothing pulled yet | Install one through ComfyUI-Manager, or download it over SSH as above |
| The disk fills up mid-session | Checkpoints and outputs add up fast on a fixed-size disk | See no space left on device |
| The link asks you to sign in again | Only the DeployGPU account that launched the deployment can open it | Sign in with that account |
The deployment failed with APP_HEALTH_TIMEOUT | ComfyUI never answered within the wait | The reserved hour is refunded automatically. Launch again |
For anything else, the troubleshooting section and the contact page are the next stops.
Frequently asked questions
Does ComfyUI on DeployGPU have its own login?
No. ComfyUI itself has no sign-in screen. The protection is the private link, which only opens for the DeployGPU account that launched the deployment.
Are any models preloaded with the ComfyUI template?
No. The template installs ComfyUI and ComfyUI-Manager only. You pull checkpoints, LoRAs and VAEs yourself after launch.
What happens to my models and outputs if I stop the instance?
They are deleted with the rest of the disk. Stop terminates the instance, so download outputs and note which models you used before you stop.
Which GPU should I pick for a FLUX or SDXL workflow?
A 48 GB card such as the RTX A6000 or L40S covers most still-image workflows at full precision. Move up to an 80 GB or larger GPU for the larger video models.
Can I install custom nodes?
Yes. ComfyUI-Manager ships with the template, so most custom nodes install from its own menu without touching SSH.
Does the ComfyUI image update itself?
No. It updates when DeployGPU rebuilds it, not automatically. If a specific ComfyUI version matters for a workflow, check the version inside the app after launch.