Kukicha / README.md
Aotsuyu's picture
Update README.md
7328dd2
|
raw
history blame
2.62 kB
metadata
license: creativeml-openrail-m
tags:
  - stable-diffusion
  - text-to-image
  - safetensors
language:
  - en

Kukicha

It's a model.

HoloKukicha is just the same model but with Vtubers merged in but it's what I mostly use so grab that.

Defaults that I use:

detailed background, masterpiece, best quality, 
Negative prompt: (low quality, worst quality:1.4), (bad anatomy), extra digit, fewer digits, (extra arms:1.2), bad hands, by (bad-artist:0.6), bad-image-v2-39000
Steps: 25, Sampler: DPM++ 2M Karras, CFG scale: 7, Seed: 1189140555, Size: 512x768, Model: HoloKukicha, Clip skip: 2, ENSD: 31337, Discard penultimate sigma: True

This is with kl-f8-anime2 vae, probably use higher cfg if trying with the anything or whatever vae, hell, mess around.

I also sometimes use 22ish cfg with:

Dynamic thresholding enable#: True, Mimic scale: 7, Threshold percentile: 96, Mimic mode: Half Cosine Up, Mimic scale minimum: 4, CFG mode: Half Cosine Up, CFG scale minimum: 4

DARKNESS

The darkKukiv1.safetensors LoRA included in the repo is my first version of a noise offset LoRA specifically for this model.

I'm working on improved versions but for now it's probably your best bet. What it does is it makes the model able to output a wider range of lighting, example:

That's with the lora at 0.6 strength and the positive prompt:

detailed background, masterpiece, best quality, 1girl, solo, hatsune miku, pov, bed, window, night, darkness, dark, blue hair, twintails, fake smile

Horror

This is an another source of noise offset so if that's an issue, disable the darkness LoRA.

And just read the readme of the LoRA.

https://huggingface.co/Aotsuyu/HorrorLora/

Big Thanks to

  • Closertodeath for being cute.
  • andite for being handsome.
  • syaimu for making a nice model that looks nice. Their model is a part of the mix. They don't know. Don't tell them.
  • NoCrypt for providing the necessary computing power.
  • Other members of the 東方Project AI community.
  • Anons responsbile for hololive/vtuber finetunes.

This feels shitpost-y

That's because it is, writing readmes is hard.

I might be arsed to improve this model card, upload fp16 and put this all on civit at some point.

That point isn't now though.