Designers, online shops and small studios in Vietnam increasingly search for "thuê GPU chạy Stable Diffusion", that is, renting a GPU to run Stable Diffusion, instead of buying a graphics card that sits idle most of the week. Renting makes sense when you generate in bursts: a product shoot for a Shopee or TikTok shop, a batch of banner ideas, a LoRA trained on your brand style. This guide covers the whole workflow on Cloudj: choosing a 12-24 GB card near Vietnam, paying from a VND top-up, reaching the web UI safely, moving images in and out efficiently, and deciding between stopping and destroying.
For a cost-first view of renting near Vietnam, read the cheapest GPU near Vietnam guide.
Pick the card by the model you use
Stable Diffusion's memory needs depend mostly on which model family you run and whether you train.
| Your work | Card to start with | VRAM |
|---|---|---|
| SD 1.5 generation, simple inpainting, upscaling | RTX 3060 | 12 GB |
| SDXL generation, ControlNet stacks, larger batches | RTX 3090 | 24 GB |
| SDXL with heavy workflows, LoRA training, quicker iterations | RTX 4090 | 24 GB |
These are general guidelines, not benchmarks. A 12 GB card can run SDXL with memory-saving options, but slowly; a 24 GB card removes most of that friction. If you mainly use SD 1.5 for product backgrounds, the 3060 is usually enough.
Find a machine near Vietnam
On Cloudj, GPUs come from a live list of independent host machines, filtered by country. Cloudj has no own data centre in Vietnam, so:
- Open the RTX 3060 deploy page, or change the model to the 3090 or 4090.
- Enter
SG, HK, VNin the Country field. You can tryVN, but machines in Vietnam are not guaranteed to be listed. - Set the disk size for checkpoints, LoRAs and outputs. SDXL checkpoints are several GB each, so count what you will load. The disk cannot be resized later.
- Pick from the results, which are sorted by total price.
If nothing suits you, add TW, JP or TH, or check back later; stock changes during the day. Since you will be clicking around a web UI, a nearby host feels noticeably smoother than one on another continent.
Install your UI and reach it safely
Cloudj does not provide a ready-made Stable Diffusion template. Start from "Ubuntu 22.04 + CUDA (SSH)" or "PyTorch + JupyterLab" and install the UI you prefer, such as AUTOMATIC1111's Stable Diffusion WebUI or ComfyUI, over SSH with git clone and pip. Follow each project's own installation instructions.
Do not expose the UI to the internet. Reach it through an SSH tunnel from your own computer, using the host and port from the instance's Connection tab:
ssh -L 7860:localhost:7860 -p <port> root@<host>Then open http://localhost:7860 in your browser. For ComfyUI, the usual port is 8188. GPU instances accept SSH keys only, and a newly added key can take up to 3 minutes to be accepted after the instance starts.
Move images in and out efficiently
Bandwidth is charged per GB downloaded and uploaded, at each machine's listed price per TB. Image work moves a lot of files, so a little care pays off.
- Upload reference images and LoRA datasets once with
rsync, which resumes if your connection drops:
rsync -avP -e "ssh -p <port>" ./references/ root@<host>:/workspace/references/- Download outputs in one batch at the end of a session, not image by image through the browser:
rsync -avP -e "ssh -p <port>" root@<host>:/workspace/outputs/ ./outputs/- Keep only keepers. Delete rejected variations on the server before downloading.
- Download checkpoints on the GPU host, not on your laptop. Pulling a model from its official source directly onto the instance is usually quicker than uploading it over your home connection.
Respect the licence of every model and LoRA you use, especially for commercial product images.
Paying in VND
Cloudj GPUs are paid from a prepaid balance in USD, which Vietnamese customers can top up by bank transfer in VND. Instances are billed per minute while running; the running price is GPU plus disk, and the price shown at deploy is locked for that instance. A deposit is held when you deploy and the unused part returns after you destroy the instance, sometimes after up to 72 hours. Live "from" prices for each card are on GPU cloud.
Stop or destroy between sessions
Many shops generate a batch of images, then stop for days. That is where the choice matters.
- Destroy when the batch is done and outputs are downloaded. The instance and disk are deleted and charges stop. Next time, deploy again and reinstall, or use a custom public Docker image that already contains your UI (custom images need a higher account level).
- Stop if you will come back within a day or two and want everything installed. You pay storage only, but the GPU is released, so starting again may wait until that GPU is free.
For most occasional users, destroying and reinstalling is cheaper than paying storage for a week. The stop vs destroy guide has a decision table, and Stable Diffusion on a GPU in Japan covers the same workflow for Japanese users.
FAQ
Which card should a small shop start with?
An RTX 3060 for SD 1.5 work; an RTX 3090 or 4090 for SDXL and LoRA training.
Is there a one-click Stable Diffusion template?
No. You install the UI yourself over SSH, or bring a custom public Docker image.
Can I open the web UI from my phone?
The safe method is an SSH tunnel from a computer. Avoid exposing the UI publicly.
Are there GPUs in Vietnam?
You can try VN, but machines in Vietnam are not guaranteed. SG and HK are the usual choices.
Start your first batch
Open the GPU deploy page, enter SG, HK, VN and choose the card that fits your model. For a full Vietnam-focused overview, see GPU cloud for Vietnam.