Draw Manga with AI at Home: LoRA, ComfyUI and ControlNet Map
Want to draw manga with AI at home using LoRA, ComfyUI and ControlNet? Start with a map of the tools before installing anything. I'm Labo, the lighthouse keeper. If you want to draw your own manga with AI at home, the first thing you need isn't a program. It's a map. The student came in this morning, laptop under her arm, and said she had installed "everything" last night and now couldn't tell what any of it was.
One island, one program. Today we only draw the map.
Some background: her father draws webtoons for a living, and he works late every night. She wants to learn AI so she can take some of that load off him.
"Kohya, LoRA, safetensors... are they all the same thing?"
Student: Labo, Kohya? LoRA? Safetensors? Same thing?
Labo: Three different layers. .safetensors is a file format. A .jpg holds a photo; a .safetensors file holds an AI model.
Student: So a LoRA is a safetensors file?
Labo: Yes, and so is the base model. The difference is size and job. The base model already knows how to draw. It's big, around 6GB. A LoRA is a small file you put on top of it, 200 to 400MB, holding only one person's drawing style. Teaching the AI your dad's style means making a LoRA.
Student: And Kohya makes the LoRA?
Labo: Right. Kohya_ss is a training program (official GitHub, open source, free). You give it a few dozen drawings, let it run for hours, and one LoRA file comes out. It only makes. It doesn't draw your pages.
GitHub publishes code in the open, so anyone can see what's inside before running it.
| Name | What it is |
|---|---|
| .safetensors | The file format all of these come in |
| Base model | The part that already draws, about 6GB |
| LoRA | A small style file, 200 to 400MB |
| Kohya | Makes LoRAs (free, your PC) |
| SwarmUI / ComfyUI | Draw with base + LoRA. Buttons / nodes (free, your PC) |
| ControlNet | Holds the composition: line art, depth |
| RunPod | Rent someone else's GPU by the hour (paid) |
"Why are there so many base models? Are they ranked?"
Student: DreamShaper, AlbedoBase, Juggernaut... is one cheap and one premium?
Labo: Not ranks, families. Three big ones. SD1.5 is older, lighter and fast. SDXL is bigger and draws better; AlbedoBase and Juggernaut are here. FLUX is bigger still. Inside one family, different names mean different people fine-tuned it on different pictures.
Student: Which family fits my dad's style?
Labo: I'd have to see his drawings. Pick by name alone and you'll pick wrong. That's a later page. Today, just the map.
"Where do I actually draw, SwarmUI or ComfyUI?"
Labo: Here's the line that keeps you from getting lost: the side that makes the LoRA and the side that uses it are different programs. Kohya makes. To use it, you load a base model plus your LoRA in a generator and say "draw this." The two main ones are SwarmUI and ComfyUI (both official GitHub, open source, free).
Student: What's the difference?
Labo: SwarmUI is buttons. Pick a model, pick a LoRA, type, generate. Easier at the start.
SwarmUI: model, LoRA, prompt, button. That's it.
Labo: ComfyUI is nodes. You connect boxes with wires and build the process yourself. Harder, but it takes precise orders like "keep these lines, only add color." To use AI in real work for your dad, you'll end up in ComfyUI. At first, just use SwarmUI to check that it works at all.
Student: Can't I just use it in Clip Studio? That's what my dad uses.
Labo: No. Clip Studio Paint is for people drawing; a LoRA is only read by AI generators. So the flow is: generate in SwarmUI or ComfyUI, bring the image into Clip Studio, and your dad fixes it. That's the shape of it in a Korean webtoon studio: the AI makes a draft, the artist finishes the page.
"Is ControlNet a LoRA option?"
Student: Line art, depth, inpaint... LoRA options? SwarmUI plugins?
Labo: Neither. A third part: ControlNet. A LoRA decides the style. ControlNet holds the shape. Feed it line art and say "color inside these lines," or feed it depth and say "keep this composition." Line art and depth are kinds of ControlNet. It goes into the generator, into ComfyUI. Some files now bundle several kinds into one, like ControlNet Union SDXL (official Hugging Face page, free).
Student: So a LoRA alone can't follow my dad's storyboards?
Labo: It can't. Style from the LoRA, composition from ControlNet. You need both for "his style, his framing," and for a consistent character across panels that sits where the storyboard puts her.
"Is CUDA something I choose?"
Labo: No. CUDA is the layer that lets an NVIDIA card do the math. If you have one, Kohya and SwarmUI use it automatically. "Local" means on your computer; CUDA means on the graphics card inside it. Not opposites.
"Where do I download models, and what's RunPod?"
Labo: No subscriptions. Kohya, SwarmUI and ComfyUI are free. Models come from Hugging Face (official models) and Civitai (models and LoRAs people have fine-tuned). Download only. A LoRA made from your dad's drawings never gets uploaded. That file is his style. So one rule from day one: the training and the LoRA file stay on your computer. That rules out every "upload your art, we'll train it" site.
RunPod rents someone else's GPU by the hour. If your home PC is too weak, you train there, take the LoRA home, and delete the rented machine. The drawings do go up, so pick a private space, not a shared one, and wipe everything after.
Student: Put it in order for me.
- Collect your dad's drawings. A few dozen. That's the textbook.
- Train in Kohya, at home or on RunPod. Out comes one LoRA.
- Download a base model. Choose the family by looking at his drawings.
- Generate in SwarmUI or ComfyUI: base + LoRA + ControlNet for composition.
- Bring it into Clip Studio. Your dad fixes it.
Student: So last night I skipped step 1 and went straight to step 2.
Labo: Exactly. Next time we check whether your computer can even do this. Have you opened the Performance tab?
Student: ...Where's the Performance tab?
FAQ
Q. Is a LoRA a safetensors file?
Yes. Both use .safetensors. The base (about 6GB) draws; a LoRA (200 to 400MB) holds a style.
Q. SwarmUI or ComfyUI?
SwarmUI is buttons, ComfyUI is nodes. Start with SwarmUI, move to ComfyUI when you need fine control. Both free.
Q. Can I use a LoRA in Clip Studio Paint?
No. Generate in SwarmUI or ComfyUI, then fix the image in Clip Studio.
What the artist did: sorted last night's installs into making (Kohya), using (SwarmUI, ComfyUI) and composition (ControlNet); set the rule that training and LoRA files stay home.
What the AI did: nothing yet. No picture was generated today.
Written October 2026. Program names and free/paid status are as of then; screens may change. The original Korean post: AI 웹툰 그리려면 어떻게 해야할까 (Korean)
Lighthouse log. Installing all night and understanding nothing by morning isn't foolishness. It's walking without a map. One map is enough.
Keep reading
- More on beginner · LoRA · ComfyUI
- All posts in Draw Manga with AI at Home


Comments
Post a Comment