First Kohya LoRA Done, Style Missing? Checking It in SwarmUI

Kohya LoRA training finished but the style is missing in SwarmUI? A run that ends does not mean the style showed up. I'm Labo, the lighthouse keeper. Today is a first LoRA training run in Kohya, from the boxes you fill in to the moment you check the result in SwarmUI, all on a computer at home. The thing that day taught me: a training run that finishes is not the same as a style that shows up.

Labo the lighthouse keeper and the student checking a first LoRA training result from Kohya, draw your own manga with AI at home

Five files on the desk, and one question: is the style in there?

The student came back the day after we picked the practice drawings and wrote the captions. Her dad draws webtoons and never gets a day off, so she's learning this to help him draw his manga with AI at home, one tired step taken off his hands at a time. Today she opened Kohya and froze.

Student: Labo, there are so many boxes. Where does everything go?

"Kohya is open. Which tab do I go to?"

Student: There are tabs across the top. Dreambooth, LoRA, Finetune...

Labo: The LoRA tab. If you set things up in one of the tabs with a similar name, you get a different kind of training. I wandered around too, then came back saying, "Right, set it up under LoRA." Inside the LoRA tab, open Training and fill in just these:

  • Base model: where the model file (.safetensors) you downloaded sits. My first run used AnyLoRA.
  • Image folder: not the 5_lbtoon01 folder itself, but the folder directly above it. Kohya looks inside for folders that start with 5_. This is the one people get wrong most.
  • Output folder and output name: where the result lands, and what it's called.
  • Epochs: how many rounds. I used 10.
  • Network Rank / Alpha: the size of the LoRA. I used 32 / 16. For a first run, that's enough.
  • Mixed precision: fp16. I once said "I can't find mixed precision" and hunted for ages. There are just that many boxes. Press Ctrl+F and search the page.
Kohya LoRA training settings for a first run, base model, image folder above 5_lbtoon01, 10 epochs, rank 32 alpha 16, fp16

Six boxes. The folder one trips up almost everyone.

Student: It won't get the style in one go, will it? We fix it bit by bit instead of starting over?

Labo: That's right. So turn on saving along the way. It leaves a file every few rounds. Later you can pick whether round 3 or round 8 looks better. Run it too long and it can actually get worse.

"I pressed Start. What came out 56 minutes later?"

Labo: I ran my first training on my home computer with 75 drawings. 56 minutes. Last time I said one file drops out. In fact there were five .safetensors files in the output folder, because the saves along the way stay there too.

Student: So this is the LoRA? Where do I open it? Does it work in Clip Studio?

Labo: Not in Clip Studio. A LoRA is only read by programs that generate pictures. And I had one more rule: nothing leaves my computer. In a Korean webtoon studio, unfinished drawings are the work itself, so I didn't want them on anyone else's server. That's how I ended up with SwarmUI.

"What about SwarmUI?"

Labo: It's button-based, so it's friendly for a first look, and it runs only on your own computer. Watch two things while installing:

  • .NET 8 installs along with it. SwarmUI runs on top of it. If it asks, say yes.
  • It asks who will use it. Choose Just Yourself On This PC, so it only opens on this machine. That's what I chose.

After that, four steps:

  1. Put the base model file in SwarmUI's Models\Stable-Diffusion folder.
  2. Put the LoRA file Kohya made in the Models\Lora folder.
  3. On the Generate screen, pick the base model, then click your LoRA in the LoRA list below to attach it.
  4. Write the trigger word at the very front of the prompt and press Generate.

Student: Don't I have to merge the LoRA into the base model?

Labo: No. You only attach it. Attached, you can swap LoRAs or turn the strength to 0.5 or 0.8. Merge them and you lose that. Later, when you're trying to keep a consistent character across panels, you'll want that freedom.

how to use a LoRA in SwarmUI, attach it to the base model instead of merging, adjust strength 0.5 or 0.8

A lens you can clip on and take off, not glue.

"Training finished. So is that the style?"

Labo: Then I attached my LoRA and made nine pictures. They were... pretty. A few had a little of the color. But the thing I'd been waiting for, the style I was teaching, wasn't there. Just an AI picture I felt I'd seen somewhere before. The 56 minutes had almost all gone for nothing.

Student: Were the settings wrong?

Labo: The settings were right. What went wrong came before them. Kohya told me "finished," and that was true. It just didn't mean the style had arrived.

One more note: when SwarmUI draws, it uses ComfyUI inside as its engine. Buttons in front, ComfyUI behind. I'd thought "ComfyUI won't run on my computer," and from that day I started moving over to it.

Student: So what came before the settings?

Labo: Next time: 56 minutes of training, so why doesn't the style come out?

FAQ

Q. Where do I set up LoRA training in Kohya?
The LoRA tab, then Training. For the image folder, choose the folder directly above the one that starts with 5_.

Q. Where do I use a LoRA made with Kohya?
In a program that generates pictures, like SwarmUI or ComfyUI. Put it in the Lora folder and attach it to a base model; don't merge it. It doesn't work in Clip Studio.

Q. What should I choose when installing SwarmUI?
Just Yourself On This PC, so it only opens on your own computer.

What the artist did: chose the base model, the folder, the rounds and the size, pressed Start, then looked at all nine pictures and said honestly, "that isn't the style."
What the AI did: trained for 56 minutes and left five files. Whether the style is in them, only the artist's eye can tell.

Official pages: Kohya_ss on GitHub (open source, free) and SwarmUI on GitHub (open source, free; the Windows install steps are on the first page). GitHub is a site where code is shared in the open, so anyone can see what's inside before running it.

Written October 2026. The numbers (75 drawings, 56 minutes, five files, Rank 32 / Alpha 16, 10 epochs) are from my first run. The illustrations and the trigger word were made new for practice; no real artwork is used. Kohya and SwarmUI update often, so box names and places may differ a little. The original Korean post: kohya 학습과 LoRA (Korean)

Lighthouse log. The lamp being lit and the ship seeing it are two different things. The first file dropping out is the most exciting moment; opening it for the first time is the most honest one.

Comments

Browse by topic

Latest posts

    See every post

    Popular posts from this blog

    Draw Manga with AI at Home: LoRA, ComfyUI and ControlNet Map

    How Many Drawings to Train a LoRA? Picks, Captions, Trigger