The model that invented Jefrie couldn't draw her again, until I trained a LoRA

Jefrie was born in a model that could never draw her again. She came out of oneObsession, an Illustrious anime checkpoint I run locally, on 8 September. For three weeks after that, the only way to get a new picture of her was to hand that one render to Nano Banana Pro as a reference. Today I trained a LoRA on my laptop, and oneObsession draws her on demand: 27 of 30 test prompts came back as her, including angles and scenes she has never been drawn in.

Here’s how she got lost, and how she came back.

She was supposed to be BloodTailor

I was trying to generate BloodTailor, my gamertag character since I was fourteen: a smooth iron mask, an impossibly wide grin, a red cloak and barbed wire around his forearms. oneObsession kept drawing him as a girl.

One of those accidents was a girl I really liked, so I stopped fighting it and started chasing her. I generated hundreds of images, tuning the prompt toward what I was seeing in my head, until this one came out:

That’s Jefrie. And I couldn’t make another one.

The prompt that made her is still BloodTailor’s prompt: 75 tags, most of it describing him (“iron mask covering his entire upper face”, “plain white button-down shirt”). She only exists because of a few tags stapled onto the end: 1girl, (little girl:1.45), snake tail. Every run was a gamble.

The tail was the worst of it, and the tail was what I cared about most. It’s supposed to weigh more than she does. No size adjective ever made it bigger. “Huge”, “giant”, “massive”, stacked weights: nothing. The only thing that worked was the phrase “coiled densely around”, and it dragged other things into the picture that I didn’t want. Reading the metadata back for this post, I noticed the phrase in her original prompt isn’t even about her tail. It’s left over from BloodTailor’s forearms: (thick rusted barbed wire coiled densely around forearms:1.5). The model heard “coiled densely” and gave it to the tail.

So she moved to Nano Banana

From then on she lived in Nano Banana Pro, with that render as the reference image. That’s where her design got finished: the half mask, the torn leather ear points, the scythe with the eye in it, the reference sheets. Everything since, including putting her in 3D and a fighting game, was built on those references, not on the model she was born in.

A thousand prompts without a LoRA

I did try to get her back into oneObsession the hard way. On the night of 10 September Claude ran an overnight experiment for me: 14 rounds, 998 renders, testing one feature at a time (mask, ears, eyes, palette, weapon, background) and scoring against the original. It measured real token counts, because SDXL quietly stops listening after 75.

It found real things. The half mask, the grey world with red as the only colour, and a two-handed grip on the scythe all became reliable. But the ears were always catgirl ears, devil horns or bows. The eyes were ordinary anime eyes. The tail was a thin devil’s tail or missing. And the features that worked alone fought each other when they were combined.

Nearly a thousand renders, and none of them is her.

The LoRA

A LoRA is a small add-on file, 85 MB here, that teaches a model one new thing. You show it captioned pictures over and over until a made-up trigger word (mine is jefr1e) means her.

I went through my library in digiKam and colour-labelled every picture I felt looked like her: 48 of them. Claude cut the reference sheets into single views, pulled nine angles out of a turntable video, and captioned all 65 images. The captions describe only what changes (the pose, the angle, the setting), so the mask, grin, tail and cape get soaked into the trigger word. Then kohya’s trainer ran on my laptop’s RTX 4070 with 8 GB: 1,480 steps, 56.6 minutes.

The first attempt sat for 19 hours and got through four images. The trainer had been quietly running on the CPU the whole time, because of a config file written during a broken first install. Claude found the line in the log that said so, fixed it, and added a guard that kills any run that lands on the CPU.

Then I gave it 30 prompts, most of them things she’d never been drawn doing:

The tail is huge in every one of them, and the prompt never mentions a tail.

Three of the 30 missed. My Grease Basket day job came out as a generic red-haired cashier, because only one training image showed it. Sleeping lost the mask and grew a fluffy tail. One close crouch let the mask slip. It also doesn’t carry over to Pony, the other model I use; that would take its own run.

Those misses are a shopping list, not a wall: five or six day-job pictures, a few relaxed poses, then train again.

What gets me is how backwards the whole thing was. oneObsession invented her by accident, and a thousand careful prompts couldn’t get her back. What finally worked was Nano Banana and the reference images drawn from that one lucky render, fed back into the model she came from.