Train your own SDXL LoRA with Kohya’s GUI on a rented GPU, in about an hour
Tutorial · Published 30 October 2026 · prices on this page are live
A LoRA is a small file that teaches Stable Diffusion a new subject or style: your product, a character, a drawing style, your own face. Training one for SDXL needs a 24 GB card for about an hour. Kohya’s GUI, the standard trainer, comes ready to run on PowerPod, so the hard part is the pictures, not the setup.
Before you rent: the pictures
- 20 to 40 images of the subject, varied in angle, lighting, background and distance. Thirty good ones beat a hundred near-duplicates.
- At least 1024 pixels on the long side; Kohya buckets other aspect ratios, so they need not be square.
- A caption per image, in a text file with the same name: a short plain description that includes a rare trigger word for the subject (“ohwx person in a red jacket on a balcony”). Kohya’s GUI can write first drafts with its BLIP or WD14 captioning tools; fix them by hand.
- One folder named
10_ohwx person: the number is how many times each image is repeated per epoch, the rest is the trigger word and the class.
Step by step
- Rent a 24 GB card with the Kohya software. Create an account, add credits (from USD 5.00, INR 200 or NPR 200), choose Kohya’s GUI (LoRA / model training) and a card with 24 GB or more from the live board. An RTX 4090 is the usual pick; an RTX 3090 costs less and takes about twice as long.
- Open the GUI and upload the folder. When the machine shows Running, open the app. Upload your image folder (zip it first), and point the LoRA tab’s image folder at it.
- Set the training. Base model: Stable Diffusion XL 1.0. Resolution 1024,1024. Network rank 16 to 32 with alpha 16. Learning rate 1e-4 for the UNet; leave the text encoders untrained for SDXL. Batch size 1 or 2, mixed precision bf16, cache latents on, gradient checkpointing on. Set epochs so the total is roughly 1,500 to 3,000 steps: with 30 images repeated 10 times that is 5 to 10 epochs. Save every epoch so you can pick the best one.
- Train. Press Start and watch the loss settle; about an hour on an RTX 4090 for 2,000 steps at 1024 pixels. Try the saved epochs as you go in the sample prompts.
- Download the LoRA. The result is a
.safetensorsfile of a few tens of megabytes in the output folder. Download it through the app’s file browser, or withscpif you attached an SSH key. - Use it. Drop the file into ComfyUI’s or Forge’s LoRA folder, load it at a strength of 0.7 to 1.0, and prompt with your trigger word. Both run ready-made on PowerPod too: the ComfyUI guide shows the way.
- Stop the machine. GPU billing ends at once. Keep the disk if you plan a second run; destroy the instance otherwise.
How long, and what it costs
A typical SDXL LoRA of about 2,000 steps, on today’s cheapest host of each card:
| For | USD | INR | NPR |
|---|---|---|---|
| On an RTX 4090, about 1 hour | USD 0.60 | INR 56 | NPR 88 |
| On an RTX 3090, about 2 hours | USD 0.60 | INR 56 | NPR 88 |
| On an RTX A6000 (48 GB), about 1.2 hours | USD 0.78 | INR 74 | NPR 116 |
| On an L40S (48 GB), about 0.8 hours | USD 0.69 | INR 64 | NPR 103 |
If the results are wrong
- Everything looks like the training photos, same pose and background: too many steps or too few varied images. Use an earlier epoch and add variety.
- The subject barely shows: too few steps, or the trigger word is a common word. Train longer, or pick a rarer token.
- Out of memory: lower the batch size to 1, make sure gradient checkpointing is on, or train at 768 pixels. A 16 GB card can manage SDXL with these settings but slowly; 24 GB is the comfortable minimum.
Questions
How many images do I need for a LoRA?
Twenty to forty varied images of the subject work for most people. More helps only if they add variety: different angles, lighting, backgrounds and distances. Near-duplicates teach the model the background instead of the subject.
Can I train an SDXL LoRA on a 16 GB card?
Yes, with batch size 1, gradient checkpointing on and a lower resolution, but it is slow and fiddly. A 24 GB card such as the RTX 3090 or RTX 4090 trains at full resolution without tricks, and the rental difference for an hour is small.
Can I train on a FLUX or Stable Diffusion 1.5 model instead?
Kohya’s GUI trains SD 1.5 LoRAs with the same steps and less memory (12 GB is enough). FLUX LoRA training is heavier: plan on a 24 GB card with memory-saving options or a 48 GB card.
Who owns the LoRA I train?
You do. It is a file on the machine you rented; download it and destroy the instance, and nothing stays behind. The base model’s licence still applies to what you generate with it.