Illustrators, manga assistants, game studios and product designers in Japan increasingly run Stable Diffusion, SDXL or Flux on rented GPUs instead of buying a workstation. A Stable Diffusion GPU cloud Japan setup on Cloudj takes four decisions: which card, which country, which interface, and how to reach it safely. This tutorial walks through each one, from choosing VRAM to opening ComfyUI in your browser over an SSH tunnel.
Step 1: choose the card by VRAM
Image models are mostly limited by VRAM. Rough, general guidance, which varies with resolution, batch size and the exact model files you use:
| Workload | Comfortable card | Notes |
|---|---|---|
| SD 1.5 and its LoRAs | RTX 3090, 24 GB | Plenty of headroom; smaller cards also work |
| SDXL with ControlNet and LoRAs | RTX 4090, 24 GB | Faster generation than the 3090 in most workflows |
| Flux and other large recent models | RTX 5090, 32 GB | Extra memory for full-size weights or bigger batches |
| Video generation experiments | 32 GB or more | Check the model's own requirements |
Many Flux workflows also run on 24 GB using reduced-precision weights. If you are unsure, start with an RTX 4090; all three cards above are consumer GPUs available at the starting account level. Live "from" prices are on each model page and on GPU cloud.
Step 2: search for a machine in Japan
Location matters here for one reason: you work interactively. Every click in ComfyUI or WebUI and every preview image travels between you and the GPU, so a nearby host feels smoother.
- Open the deploy page with the RTX 4090 preselected, or pick another model tile.
- Type
JPin the Country field. - If a machine is listed, check reliability, max duration and download speed. Download speed matters because checkpoints are several gigabytes each.
- If nothing is listed, add neighbours:
JP, KR, TW. Stock changes during the day, so you can also try again later.
Renting a GPU in Japan explains the JP search and fallbacks in more depth, and the RTX 4090 in Japan focuses on that card.
Step 3: deploy with the right template and disk
- Template: PyTorch + JupyterLab gives you a working PyTorch install, which saves time with ComfyUI or WebUI. Ubuntu 22.04 + CUDA works too if you prefer to install everything yourself.
- Disk: checkpoints, LoRAs, ControlNet models and outputs add up quickly. Choose generously, because disk size cannot be changed after deploy.
- SSH key: select at least one; password login is disabled. A newly added key can take up to 3 minutes to be accepted after the instance starts.
Note that Cloudj does not ship ComfyUI or WebUI as templates. You install the interface yourself, which also means you control its version and extensions.
Step 4: install ComfyUI over SSH
Connect with the host and port shown in the instance's Connection tab. The port is not always 22.
ssh -p PORT root@HOST
nvidia-smi
git clone https://github.com/comfyanonymous/ComfyUI.git
cd ComfyUI
pip install -r requirements.txt
tmux new -s comfy
python main.py --listen 127.0.0.1 --port 8188Running inside tmux keeps ComfyUI alive if your SSH session drops. Detach with Ctrl+b then d. Stable Diffusion WebUI installs in a similar way from its own repository; follow its instructions for Linux.
Download checkpoints into ComfyUI/models/checkpoints/, LoRAs into models/loras/, and so on. Pull them directly on the GPU host rather than uploading from your PC: the host's download speed is usually far better than a home upload.
Step 5: open the interface through an SSH tunnel
Do not expose ComfyUI to the internet. Instead, forward its port to your own computer:
ssh -p PORT -L 8188:localhost:8188 root@HOSTLeave that terminal open and browse to http://localhost:8188. The interface runs on the GPU in Japan, but only you can reach it. How to SSH into a GPU instance covers keys, ports and tunnels in more detail, and running ComfyUI on a GPU in Europe shows the same pattern for European users.
Getting your images home
Generated images pile up in ComfyUI/output/. Copy them to your machine regularly:
rsync -avP -e "ssh -p PORT" root@HOST:/root/ComfyUI/output/ ./outputs/Adjust the path to wherever you cloned ComfyUI. Do this before you destroy the instance and before the machine's "available until" date, when the instance and its disk are removed.
Saving money between sessions
GPU instances are billed per minute while running. When you finish for the day you have two options:
- Stop: keeps your disk and charges storage only, but releases the GPU. Someone else may rent it, and restarting waits until it is free.
- Destroy: deletes everything and stops all charges. Keep your workflow JSON files and a list of model downloads so you can rebuild quickly on a new machine.
Licences and responsibility
Check the licence of every checkpoint and LoRA you use, especially for commercial work; some popular models restrict commercial use. Respect the terms of the sites you download from and the rights of artists whose styles you work with.
FAQ
Which GPU is best for SDXL?
A 24 GB card such as the RTX 4090 is comfortable for SDXL with ControlNet and LoRAs. The RTX 3090 also works well.
Does Cloudj preinstall ComfyUI?
No. Install ComfyUI or WebUI yourself over SSH, starting from the PyTorch or Ubuntu + CUDA template.
Can I open ComfyUI directly from a public URL?
Use an SSH tunnel instead. It keeps the interface private and needs no extra setup.
What if no GPU is listed in Japan?
Add KR and TW, try later, or choose another card with enough VRAM.
Next step
Open the RTX 4090 deploy page, set the Country field to JP, and have ComfyUI running in your browser within the hour.