Sketch to AI Manga: Does ControlNet Scribble Keep Lines?

Does ControlNet scribble keep the lines of your own sketch when turning it into AI manga? Scribble worked better for me than lineart mode. I'm Labo, the lighthouse keeper. Today's page is about how to turn your own sketch into manga with AI: putting a sketch into ControlNet scribble, and why I tried lineart mode first and came back to scribble. What I learned: half of a drawing style comes from the sketch. When the sketch is already the artist's, I don't force a style LoRA on top.

Turn your own sketch into manga with AI, ControlNet scribble keeps the artist's lines, draw your own manga at home

The picture I used to describe in words, I now show in lines.

"If the prompt isn't enough, does a sketch change the style?"

Student: Labo, last time you said a phrase that wasn't in the captions won't work in the prompt. So prompts can do less than I thought?

Labo: For me, yes. The LoRA only understands the words it learned, and even then I couldn't write "this person, this angle, this spot" as exactly as I wanted.

Student: Webtoon panels have a fixed layout. You'd have to write all that out?

Labo: Exactly. The prompt alone wasn't getting there, so I decided to put the sketch in myself. Show the layout in lines instead of words. That's drawing your own manga with AI at home: you still draw, and the AI follows.

Student: What kind of sketch? Rough or finished?

Labo: I ran two, because sketches look different at different stages of work:

  • A sketch with shading, shadows painted over the lines
  • A sketch with black-and-white lines only, no shading

The shaded one is for next time. Today is how the sketch goes in.

"How does ControlNet follow a sketch?"

Student: So you lay the sketch underneath and the AI draws over it?

Labo: Not quite. This is ControlNet, the "layout part" from the first map post. A LoRA teaches style; ControlNet holds "lines here, person here." Separate parts.

ControlNet has modes that decide what it reads from your sketch:

  • scribble reads the flow of bold lines, the big structure of what's where
  • lineart reads the lines of a line drawing, as the name says
  • canny finds edges where brightness changes sharply and reads them as outlines

I did this in ComfyUI. The nodes that prepare a sketch for each mode come in ControlNet-Aux, an extension I installed when I moved to ComfyUI.

"Isn't lineart the obvious mode for line art?"

Student: But webtoon sketches are lines. The name is perfect!

Labo: I thought exactly that, so I ran lineart first. It came out strange. I only wanted the sketch's structure: who stands where, which way the body faces. But lineart mode copied the color too. It grabbed things I never asked for.

Student: You told it to follow lines and it followed the color?

Labo: Yes. So I went back to scribble. It reads only the big flow, holding the structure and leaving the AI room to paint the rest. Why lineart pulled in color isn't in my records.

ControlNet scribble vs lineart vs canny on the same sketch, scribble keeps structure only, lineart copied color

Same sketch, different mode, different things held. An explanation drawing, not a real render.

Labo: So don't pick by the mode's name. Pick by what you want it to hold. If it's structure only, start with scribble.

"What strength should I use?"

Labo: Here are my settings from my notes, in the ControlNet node and the KSampler node:

NodeBoxMy value
ControlNetmodescribble
ControlNetstrength0.85
ControlNetstart / end0 / 0.9
KSamplersteps30
KSamplerCFG6.5
KSamplersampler / schedulerdpmpp_2m / karras
KSamplerdenoise1.0
  • Strength: how tightly it sticks to the sketch. Closer to 1 is tighter.
  • Start / end: if making the picture runs from 0 to 1, when ControlNet holds on. 0 to 0.9 means almost the whole way, letting go only at the very end.
  • Steps 30: how many passes to clear the noise. On my RTX 3050, SDXL at 30 steps ran about 10 seconds a step, 6 to 7 minutes a picture.
  • CFG 6.5: how stubbornly it follows the prompt.
  • dpmpp_2m, karras: how the noise is cleared, and its pace. Fine as is at first.
  • Denoise 1.0: draw new, with no base picture. ControlNet supplies the structure.
ControlNet scribble strength 0.85 start 0 end 0.9 and KSampler 30 steps CFG 6.5 dpmpp_2m karras denoise 1.0, sketch to manga settings in ComfyUI

My values on one board. The second dial is the 0.9 to 0.95 from my September note.

Student: So 0.85 is the answer?

Labo: Honestly, my records have two numbers. My settings note says scribble 0.85. My September note says to keep strength at 0.9 to 0.95 to keep the artist's lines alive. I don't merge them; a middle value would be a number I never tried. The September note is clear: lower the strength and the artist's lines blur.

Student: Then what do I do?

Labo: Render the same sketch with the same seed at 0.85 and at 0.9 to 0.95, side by side. Whichever keeps the lines alive is that sketch's answer. One note: the ControlNet 1.0 from when my epochs 4 and 6 looked the same was a different job.

"Half the style comes from the sketch?"

Student: Sketch for structure, LoRA for style, perfect, right? You worked so hard on those LoRAs.

Labo: I planned to use both. But in October I set one rule: the sketch already holds the artist's style, so I don't force a style LoRA in. Every line an artist draws is already their hand: face proportions, where a line breaks, how a body bends. If ControlNet holds those lines, the artist is already in the picture. I'm not dropping LoRAs, just not forcing one onto a panel with a sketch.

Student: I want to try with my dad's sketches. But what happened to the shaded one?

Labo: I put the shading in carefully, and it came out far too flat. That's next time.

FAQ

Q. ControlNet scribble or lineart?
Scribble follows the structure. My lineart run copied color too, so I went back to scribble.

Q. What ControlNet strength should I use for a sketch?
My settings note says scribble 0.85, start 0, end 0.9. My September note says 0.9 to 0.95 keeps the artist's lines. Render both with the same seed and compare.

Q. Do I still need a style LoRA if I use my own sketch?
Since October I don't force one in. A sketch the artist drew already carries the style.

What the artist did: drew the layout and lines as sketches (one shaded, one lines only), chose scribble, set the strength, and decided not to force a style LoRA on top.
What the AI did: ControlNet held the sketch's structure; the model drew and painted it new from there (denoise 1.0).

Official pages: ComfyUI on GitHub and ControlNet-Aux on GitHub (both open source, free). GitHub is a site where code is shared in the open, so anyone can see what's inside before running it.

Written October 2026 from Labo's own records. Values are from my RTX 3050 8GB and my work; each drawing may differ. ComfyUI and its extensions may change. The original Korean post: ControlNet 스케치를 넣으면 AI 그림체가 달라져요? (Korean)

Lighthouse log. Half the style comes from the sketch. When the route is already drawn by hand on the chart, you don't borrow someone else's light.

Comments

Browse by topic

Latest posts

    See every post

    Popular posts from this blog

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

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

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