Kohya SS is a web interface for the sd-scripts code that trains your own LoRA, DreamBooth and fully fine-tuned models for SDXL, SD 1.5 and FLUX. On gpu.nz it starts from one template: rent a card, open the interface and upload your dataset. You pay by the minute while the server runs, and a stopped server uses no GPU minutes while its files stay on the disk.
In short
Kohya SS runs on cards with 16 GB or more. The most affordable suitable card right now is the RTX A4000 16 GB from 13 ₽ per hour, billed by the minute.
Updated 13 October 2026
LoRA on SDXL and SD 1.5 needs 16 GB of video memory or more: an RTX 4080 16 GB handles it, while an RTX 4090 24 GB gives room for large resolutions and trains noticeably faster. For FLUX take 24 GB or more: on 16 GB FLUX is possible only with fp8 weights and block swapping, and that is much slower. Full fine-tuning and DreamBooth with larger batches are best on an RTX A6000 48 GB.
Prices and availability come from the live catalog and update automatically.
For a character LoRA, 15–30 photos with different angles and lighting are usually enough. Each photo has a text file with a caption of the same name next to it. Kohya expects a folder named like 10_ohwx man: the number before the underscore is how many times each photo in it repeats per epoch, and ohwx is a rare trigger word that you use later to call the model.
train/
img/
10_ohwx man/
01.jpg
01.txt # ohwx man, portrait photo, soft daylight
02.jpg
02.txt # ohwx man, full body, city streetUpload the whole folder with scp. Files in /workspace are kept on the server's disk:
scp -r ./train <user>@<address>:/workspace/ -P <port>Steps are counted as photos × repeats × epochs ÷ batch size. For example, 20 photos with 10 repeats and 10 epochs at batch size 1 give 2,000 steps.
| Photos | Repeats | Epochs | Batch | Total steps |
|---|---|---|---|---|
| 20 | 10 | 10 | 1 | 2,000 |
| 30 | 10 | 8 | 2 | 1,200 |
| 15 | 20 | 10 | 1 | 3,000 |
Time depends on the card, the resolution and the batch size. For SDXL on an RTX 4090, a run of about 2,000 steps usually takes between half an hour and an hour. A short test gives a closer figure: run 50 steps, time one step in seconds and multiply by the total step count. More in the guide which GPU to pick for LoRA.
Set the Output folder to /workspace/train/model. Kohya saves intermediate LoRA versions by epoch and the final .safetensors file there. Download it through JupyterLab: right-click the file and choose Download, or run this command from your computer:
scp -P <port> <user>@<address>:/workspace/train/model/*.safetensors ./A LoRA file is usually tens to hundreds of megabytes, so it is easy to move into the LoRA folder of your generation interface, such as Forge or ComfyUI.