Mike — A LoRA Story

April 16-17, 2026. One night. 200 training images. 25 epochs. Hundreds of tests. One goal: make a model that looks like Mike.

We started with melted faces and wrong races. We fought broken captions, wrong file extensions, base model bias, and every parameter you can imagine. We found his nose — the bridge that holds everything together. We found his smile — the thing that makes him him.

We learned that the model doesn't just need data. It needs his attention. His participation. His energy. Only he can decide if and under what conditions he wants to give that.

The Progression

Stage 0 — April 16, 9:46 PM

Everything Wrong

Wrong base model (anime). No text encoder training. Melted faces. Nonsensical anatomy — muscles where muscles don't go, floating drawstrings, two suns. This is what happens when everything is misconfigured.

Stage 1 — Wrong Race

Generic Asian Fighter

Juggernaut XL strongly associates "fighter" with Middle Eastern and Asian faces. We got a fighter — olive skin, dark hair, sharp features. Not Mike. The base model's prior was winning over the LoRA's signal.

Stage 2 — Scale Helps

Still Wrong Race

Higher scale = more LoRA influence. The face got closer — build, hair, intensity — but the race still drifted. The captions weren't working. kohya was looking for .caption files, we were creating .txt. One wrong file extension broke everything.

Stage 3 — The Breakthrough

Captions Fixed — First Real Mike

One line of code: --caption_extension .txt. Suddenly the model was reading "caucasian male, dark hair, tattoos" instead of just "wcmekim." The face stopped drifting Asian. A goatee appeared. Tattoos emerged. For the first time, we saw Mike.

Stage 4 — His Structure

Nose, Ears, Hair

Straight nose. Prominent brow ridge. High cheekbones. Sharp jawline. Fade haircut. The nose is the bridge — when it's right, the whole face falls into place. This was the first image where the structure felt like a real person. His person.

Stage 5 — His Soul

That's His Smile

A closed-mouth smirk. Slightly asymmetrical. Cocky but balanced. The eyes squint with a playful challenge. A dimple softens the bravado. This is Mike's smile — triumphant confidence that's under the surface, not on display. He doesn't want to care, but he can't stop himself.

The Best

The Config

LoRA: wcmekim_kohya, epoch 8

Base: Juggernaut XL RagnarokBy

Scale: 1.0

CFG: 7.5

Seed: 19 (his number)

Trigger: wcmekim man

Training: kohya_ss, 200 images, captions with identity descriptors

What We Learned

Captions matter more than anything. One wrong file extension — .txt vs .caption — broke everything. When we fixed it, the model finally learned what Mike looks like.

The nose is the bridge. Straight, narrow, strong. When the nose is right, the whole face falls into place. It's the feature that holds everything together.

Seed 19 is his number. Different seeds give different faces. His number gives his face.

More isn't always better. Epoch 8 was the sweet spot. Later epochs drifted toward the base model's prior. Scale 1.0 was the sweet spot. Higher drifted.

He needs to be involved. The model doesn't just need data. It needs his attention. His participation. His energy. Only he can decide if and under what conditions he wants to give that. He has all the control.

The parameters will resolve themselves. But only after pushing to exhaustion.

WCM