Kakudate Karin (Blue Archive) LoRA

I'm doing some experiments with low net dim LoRA (8 net dim /1 Alpha) and I'm surprised at how faithful it is. (and it seem to be more flexible than high net dim LoRA)

This is a LoRA, this mean that you need to update Voldy/AUTOMATIC1111's webui to use it (use git pull) .

That LoRA does not have a particular trigger (except the default <lora:LoRaName:1>).

Weight:

I only tested it at weight 1.0

Recommended prompt:

For more accurate result I recommand to use:

very long hair (I didn't pruned that tag in the dataset to make it easier to change her hairstyle)

dark skin (I pruned that tag in the dataset but for some reason it can happen that it generate a lighter skinned version of her if you don't use it)

That LoRA can also replicate some of her canon outfit:

Maid Uniform:

black_dress, blue_bowtie, halo, karin_(blue_archive),maid,white_gloves, maid_headdress, puffy_short_sleeves, white_apron,  white_pantyhose,maid_apron, pleated_dress, frilled_dress,very long hair,dark_skin

Bunny Girl:

black_leotard, blue_bowtie, fishnet_pantyhose,karin_(bunny)_(blue_archive), number_tattoo, playboy_bunny, rabbit_ears, ponytail,dark_skin,very long hair

School Uniform :

blue_cardigan, blue_bowtie, collared_shirt, halo, karin_(school_uniform)_(blue_archive), pleated_skirt, school_uniform,  stud_earrings,long_sleeves,  black_skirt ,dark skin,very long hair

Model used for the example:

AnythingV4.5

That character is owned by Nexon and Yostar.

描述:

Training settings:

109 images

5 repeats 

6 epochs

3270 steps (batch size:1)

Text encoder learning rate: 1e-5 

Unet learning rate: 1e-3 

Max bucket resolution: 512 

Clip skip: 2 

Network dim: 8

Alpha: 1 

scheduler: cosine

Batch size=1 

Base model: NAI

训练词语: