ComfyUI on cloud GPUs, without rebuilding it every session
Rent the GPU you need, deploy the ComfyUI template, and keep the studio you built on it — custom nodes, model weights, LoRAs and all — instead of reinstalling from scratch the next time you sit down to work.
The part nobody prices in
A rented GPU is cheap and easy. The environment on top of it is neither. A working ComfyUI install is a long tail of custom nodes at specific commits, a Python environment that resolved cleanly once, tens or hundreds of gigabytes of checkpoints and LoRAs, and a settings graph you tuned by hand over weeks.
Rent a box, and none of that comes with you. The GPU is stateless by design, so every fresh instance starts at the same place: re-clone the nodes, re-resolve the dependencies, re-download the models, rebuild the graph. That work costs you real time and real GPU-hours, and you pay it again on the next session — and again the moment you switch to a different card or a different provider.
Three ways a ComfyUI setup disappears
Custom nodes reinstall
A node manager records repo URLs and commit hashes, not the files you actually ran. A rebuild re-clones and re-resolves, and upstream has moved since.
Models re-download
Checkpoints and LoRAs are the biggest thing on the box and the first thing a fresh instance is missing. You pay for GPU time while it pulls.
The studio resets
Switch cards or providers to chase price or availability and you are back at an empty ComfyUI, rebuilding a setup you already had working.
How Aquanode answers it
1. A ComfyUI template, so the first hour isn't setup
Choose ComfyUI from the Template dropdown when you deploy. Aquanode provisions the box, installs ComfyUI, and starts it once the VM is ready — you don't configure anything on the box yourself. When the deployment reaches Ready, the Overview tab gives you the ComfyUI URL plus a generated username and password.
2. Snapshots that capture files, not a list of repositories
Point a snapshot at an absolute path — your custom_nodes directory, your models directory, the whole ComfyUI tree — and Aquanode captures what is actually on disk there. Restoring puts those files back at the same paths, so an installed node comes back installed rather than being re-cloned and re-resolved against whatever upstream looks like today. Take one on demand, or set an interval and let it run automatically. Snapshots are incremental, so the second one only carries what changed.
3. A backup that isn't tied to the box that made it
You can restore to any saved backup on any GPU, not just the one it was created on. That is what lets you follow price and availability instead of being anchored to one provider by the cost of rebuilding. Being straight about it: moving to another provider is three steps today, not one button. Snapshot the current VM, launch a new VM on the provider or GPU you want, then restore the snapshot onto it. The steps are documented and each one is a click, but it is a flow, not a single action.
Network volumes are a different tool
Aquanode also offers network volumes, and they're genuinely useful — but they solve a narrower problem. A volume can only be attached to instances on the same provider and in the same region where it was created, and they're currently available on DataCrunch and Hyperstack. If your ComfyUI work stays on one provider in one region, a volume is a fine home for a large model library. If you want to move, that's what snapshots are for.
Start in four steps
- 1Pick a GPU from the marketplace and select ComfyUI in the Template dropdown, then deploy.
- 2When the deployment is Ready, open the Overview tab and click Open in the ComfyUI section — the username and password are shown next to it.
- 3Install the custom nodes you use and pull down the checkpoints and LoRAs you work with, the same way you would on any box.
- 4Take a snapshot of the paths that hold that work. Next session, deploy a GPU and restore it — your nodes and models come back at the paths ComfyUI expects.
Go deeper
We've written about each of these failures in detail — what breaks, why the common workarounds don't hold, and what actually fixes it.
Why your ComfyUI custom nodes reinstall every session
What ComfyUI-Manager's snapshot actually records — repo URLs and commit hashes, not the installed files — and why a restore can still come back broken.
How to stop re-downloading your models on every new box
The usual workarounds for a model library, where each one breaks, and why a network volume alone doesn't survive a change of provider.
Run ComfyUI on any cloud GPU: a 5090 today, an H100 tomorrow
Moving ComfyUI between providers is a version-resolution problem before it is a storage problem. What actually breaks on the new box.
Your ComfyUI studio shouldn't reset every time you switch GPUs
Cloud GPUs stayed stateless while ComfyUI setups became stateful. Why that gap lands on you every single session.
The best RunComfy alternative: 6 ways to run ComfyUI on cloud GPUs
Managed ComfyUI clouds, raw GPU pods, buying your own card, and portable studios — priced and compared side by side.
ComfyUI on Aquanode — FAQ
Do my ComfyUI custom nodes survive between sessions?
They do if you snapshot the directory they live in. Aquanode's snapshots capture the actual files at an absolute path you choose, so restoring brings the installed node back as it was on disk — it does not re-clone the node's repository and re-resolve its dependencies. That distinction is why a restored environment starts working immediately instead of failing on a dependency that has moved on since you installed it.
Can I restore my ComfyUI setup onto a different GPU?
Yes. A backup is not tied to the deployment that created it, so you can restore it onto a different GPU or a different provider. Today that is a three-step flow rather than a single button: take a snapshot of your current VM, launch a new VM on the GPU you want, then restore the snapshot onto it.
Do I have to install ComfyUI myself?
No. Pick ComfyUI from the Template dropdown when you deploy and Aquanode provisions the box, installs ComfyUI, and starts it for you. Once the deployment is Ready, the Overview tab shows the ComfyUI URL along with a generated username and password.
What about my model library — do I re-download it every time?
Only if you don't snapshot it. Model weights, LoRAs, and checkpoints are just files on disk, so a snapshot of the directory holding them restores them at their original paths on the new box. Snapshots are incremental, so repeated snapshots of a large model directory only carry what changed.
Is a network volume the same thing as a snapshot?
No, and the difference matters if you move around. A network volume can only be attached to instances on the same provider and in the same region where it was created, and volumes are currently available on DataCrunch and Hyperstack. A snapshot has no such constraint — it restores onto any GPU you deploy.
How is a ComfyUI deployment billed?
A ComfyUI deployment is a running GPU instance and it keeps billing until you close the deployment from the console. Aquanode compares live prices across independent GPU providers, so you pick the rate you want before you deploy.
