Kaoru [Femboy], otoko-no-ko

Version SD 1.5

Due to recent CivitAI's recent buzz update, i started thinking about doing my gens locally, but since my potato pc can't run full SDXL models, that means i can't use SDXL/Illustrious LoRA anymore (it doesn't even load in my webui lora tab)

This version was trained specifically for accessibility, lighter and faster generation especially for older or lower spec setups lke mine. Don't expect any major visual improvement, i reused the same dataset from v1.0, just using different base model and training param.

Also.. for some reason during training, all 10 epochs looked bad and had face distortion, but everything else looks perfect. When i run it locally, there’s no distortion at all, so if that happens on your end, i apologize.

My local setup :

Base model (the one i'm using, not the one used for dataset) : MeinaHentai V5 - final

VAE : using the one from MeinaHentai
CFG : 5-7
Sampler : DPM++ 2 SDE with Karras schedule type or euler a with automatic schedule type

Steps : 20-30

Training : SD 1.5 base, 64 networks dim with 32 alpha, 6 batch 21 repeats, 1 clip skip, no flip aug

Why I made this

The reason i made this is because at some times i was randomly generating a femboys, and out of many this the one i really really likes, so i wanna make my own models based of him just so i could only use few words to generate him rather than describing all of the details. which did not work as i intended, i mean this models was just like a test run (i put 20 images for a dataset lmao which is the reason why its not consistent).

Still working on v2 with much more images on dataset + better tagging so it will generate more consistent kaoru look i'm looking for.

Also i named him kaoru which mean smell cuz i wanna smell his armpits, i can't find any characters that looked like this so i apologize if there is a character that look like him, i don't wanna get accused of stealing characters.

This is the first time i'm creating a models so i'm sorry if the quality is bad, recommended strength is ~0.8-1.0,

• General :

1boy, male focus, kaoru, otoko-no-ko, femboy, crossdresser boy, choker, jewelry, bracelet, earrings, necklace, black hair, short hair, wolfcut, one side longer, gradient hair, hair between eyes, parted lips, two-tone hair, red hair, flat chest, red eyes, pink eyes, 

sometimes "flat chest" won't work cuz civit will think that it violates TOS (maybe cuz there's 1boy prompts?) but you could just remove it and add these negative-

• Negative :

blue hair, blue highlight, pink hair, white hair, breast, small breast, 

my usual setup :

CFG Scale : 5

Sampler : Euler a

Steps : 29

v1.1 change :

fix the bluish hair problem (i think), if you didn't mind that then just use v1.0, also add pink hair in negative prompts cuz that problem is still there, maybe i need to retrain with new dataset to fix it.

just use v1.0 and add the problem on negative prompts, like : pink hair, blue highlight, blue hair,

this version just changed him into a girl