[ANIMA style LoRA] fi4rfs - Flat InCase for @RustyFidgetSpinner

It's all in the title! A flat InCase inspired style to generate using ANIMA the elf-girls from @RustyFidgetSpinner

Update! What? A V2 already? Yes! This version is a slight alteration of V1, just a little bit less flat. I felt it matched closer the original style. You can choose either to your liking <3

Usage

Works best with base ANIMA, weight 0.8~1.0. No trigger, no specific things to add to prompt. Check showcases ;-)

The Story™

What can i say beside... well, since i have moved on to ANIMA, i am always sad to be unable to tests some characters which are made for Illustrious.

So, with that in mind and Binky in particular, i told myself "ok, let's go".

After a quick research, i found that the "standard" style that i liked on most of the Mandratha girls is in fact a "Flat color InCase" style. I may not be all for "copying" artists, but this felt a bit different if it could packaged with the flat style baked in.

So, with this idea in mind, i made around 350 pictures and did two pass of training: 10 epochs with 250 pictures at resolution 1216 and 12 epochs with 100 pictures at resolution 1536.

To stick to the original @RustyFidgetSpinner style, i used:

• WAI 14

• PPW 8_2

• Merged those two as a single model (because PPW is an almost 1GB LoRA)

• Sticked to the 29 steps - Euler A - Normal used

• Leveraged these base Positive and Negative from Elf-Sensei

HDR, 8K, high contrast, masterpiece, best quality, amazing quality, very aesthetic, newest, solo,
(artist:incase:0.8), (artist:tsuaii:0.1), (flat colors:1.6),
<my prompt>

(frame, border),
bad quality, worst quality, worst detail, sketch, four fingers, three fingers bad hands,
worst quality, low quality, displeasing, text, watermark,
bad anatomy, ((artist name, signature))

And here we are! After merging epochs 8, 9, 11 and 12 together of my 64/32 LoRA, i did a quick "compression" pass to reduce its size (with a CUTOFF at 99%).

And of course, no trigger words ^^

NB: the only "new" trick i did this time is that after cropping some pictures manually, i did a pass of a Python script to ensure all pictures dimension where a multiple of 32 => this allowed me to have a very low Aspect Ratio error when the latent were made.

A few ugly lines of python! (this takes all PNG pictures, convert them to jpg at 100 quality after shuffling them, with the new crop32 added. I can do that on a full folder of manually selected and cropped pictures before going for tagging).

from tqdm import tqdm
from PIL import Image
from random import shuffle
from os import remove
from glob import glob

liste = glob("*.png")
shuffle(liste)

def crop32(im):
  x = im.size[0] % 32
  y = im.size[1] % 32
  return im.crop((int(x/2),int(y/2),im.size[0]-x+int(x/2),im.size[1]-y+int(y/2)))

for i, f in enumerate(tqdm(liste)):
  im = Image.open(f).convert("RGB")
  im = crop32(im)
  im.save(f"{i:03d}.jpg", format="JPEG", subsampling=0, quality=100)
  remove(f)