LoRA Breaking the Body? A Separate Anatomy Dataset First
LoRA keeps breaking the body of your character? Train a separate anatomy dataset first, so the body is learned apart from the outfit style. I'm Labo, the lighthouse keeper. Today's page is about what I did when my LoRA kept breaking the body: I split the training into two sets, one of outfit drawings for the style and one where only the structure of the body shows, and taught the body on its own. If you want a consistent character across panels, this is the step that finally worked for me. What I took from it: I think the body comes before the clothes.
Put the outfits aside for a day and look at the frame underneath.
Last time I ended with "teach the body separately," and the student said it the moment she opened the door. Her dad draws webtoons with no days off, and she's learning to draw manga with AI at home to help him.
"If a LoRA keeps breaking the body, you teach the body separately?"
Student: Labo, what does "separately" mean? You make another LoRA?
Labo: First, why. Last time, the 55-image LoRA did close-ups well but failed full-body shots twice. The same woman in different outfits didn't read as one person, and the body didn't come out right.
Student: And your guess was that the outfits were so varied it never learned the body.
Labo: Right. So I went the other way. I taught the body separately, with pictures where no outfit gets in the way and only the structure of the body shows. Like figure drawing in art class.
Student: Oh, like when the teacher puts a wooden mannequin on the desk and we all draw it?
Labo: That feeling exactly. When you learn to draw, you don't start with folds in a sleeve. You look at how the head, neck, shoulders, arms, waist and legs connect, and the clothes go on top. I tried that same order with the AI.
"What kind of body pictures go in?"
Labo: One rule: the structure of the body is visible, not hidden under clothes. Put outfit pictures in and it learns the outfits along with everything else, so in this set I left the outfits out.
Student: That's your "only the thing you're teaching should show" rule again.
Labo: It's been coming back since my first LoRA. In the style set, the style shows. In the body set, the body shows. Pack too much into one image and the AI doesn't know what to learn.
Labo: I can't show the real training pictures here, so picture a wooden mannequin in different poses.
"How many style pictures and body pictures?"
Labo: In numbers:
- Style dataset (outfit drawings): 50 images
- Body dataset: 79 images
- Both at 7 repeats
- Base model: wai Illustrious (v170)
Student: More body pictures than style pictures?
Labo: Yes, 79 on the body side, 50 on the style side. And I kept them apart, not mixed in one bundle.
Student: Repeats 7? Didn't you say 5 before?
Labo: Good memory. The number at the front of the training folder name is the repeat count. In the example folder C:\lora_example\5_lbtoon01, the 5 means each image is seen five times per epoch. The 55-image LoRA used 5. This time both sets were set to 7.
Two crates, kept apart: the style set and the body set. Both at 7 repeats, on the same base.
"Why switch from Albedo to wai?"
Student: Wait. The base is wai? You downloaded Albedo on RunPod.
Labo: I did. I first trained on Albedo (AlbedoBase XL). Then I moved to wai Illustrious because of the body and the line work. Both are SDXL family. But the rule I set at the same time matters more than the switch.
Labo: Generate on the base model you trained on. A LoRA trained on wai gets attached to wai. I don't attach it to Albedo.
Student: If both are SDXL, can't it go on either? The files look the same.
Labo: It will load. But a LoRA doesn't draw by itself. It rides on a base model and learns how much to change what that model already draws. Take the "change it this much" it learned on wai and put it on Albedo, and the starting point is different, so you don't get the same result.
Student: Like the same makeup looking different on a different face.
Labo: Good comparison. Whenever you make or download a LoRA, write down which base it was trained on.
Learned on wai, drawn on wai. The other gate stays closed.
"So the body comes before the clothes?"
Student: So how did it go, teaching the body on its own?
Labo: It's working well. 50 style images and 79 body images kept apart, trained on wai, generated on wai.
Student: Wow, you finally said "it worked."
Labo: I know. Until now it's been wrong pictures and failed full-body shots. What stayed with me: I think the body comes before the clothes. Teach the structure first, like figure drawing, and the clothes go on top. And generate on the model you trained on.
Student: So is this the end?
Labo: Not quite. Once the pictures were split, something else needed care: the short text attached to each image, the caption. However well you pick the pictures, the AI reads that text and learns from it too. Next time I'll tell you about handing those captions to three AIs, Claude, ChatGPT and Grok, and checking them against each other.
Student: Three AIs? Do they all write the same thing?
Labo: If they did, there wouldn't be a next post.
FAQ
Q. Can I teach a LoRA anatomy separately?
I taught the body with a separate set of pictures where only its structure and pose show, like figure drawing in art class. It's working well.
Q. How many style images and body images did you use?
50 outfit drawings for the style and 79 for the body, both at 7 repeats, on wai Illustrious (v170).
Q. Can I attach a LoRA trained on one SDXL base to another?
It will load, but I don't. My rule is "generate on the base you trained on," because a LoRA is a difference learned on top of that base.
What the artist did: picked 50 outfit drawings and 79 body-structure pictures, kept them apart, moved the base from Albedo to wai Illustrious, and kept the rule "generate on the base you trained on."
What the AI did: trained both sets at 7 repeats on wai Illustrious (v170), then drew with the LoRA attached to wai.
Official pages: Kohya_ss on GitHub and ComfyUI on GitHub (both open source, free; GitHub shares code in the open, so anyone can see what's inside before running it), Hugging Face (where base models are downloaded), Claude (Anthropic), ChatGPT (OpenAI), Grok (xAI).
Written October 2026 from Labo's own records. Image counts, repeats and the base model version (wai Illustrious v170) are from my run. Program screens and model versions may change. The original Korean post: 로라로 인체가 자꾸 틀어질 때, 몸부터 가르치면 돼요? (Korean)
Lighthouse log. A lighthouse is built of stone before it is ever lit. I think drawings are the same: the body before the clothes. And the lamp is lit on the very stones it was built on.
Keep reading
- Previous post: Why a Consistent AI Character Changes With Every Outfit
- More on LoRA · SDXL
- All posts in Draw Manga with AI at Home



Comments
Post a Comment