After extensive testing, it is sure that Pony model cannot accept enough content within the acceptable time limit for training.
Aurora is achieve sufficient results after 44k images overtrain,but under overtrain almost lora train on Pony cannot work well on aurora.
Tests with 89k and 200k images almost did not produce any effective results.
Conversely, only a few hundred datasets yielded good results.
In conclusion, aurora can no longer be updated base pony.After all,if a larger training volume is required, why not skip Pony and directly fine-tuning base SDXL 1.0?
---------------------------------------
The performance of Kohaku XL Delta is over my expectations, and I'm not a Furry.
So this model will stop all update base the Pony XL.
please wait new kohaku XL.2024-03-03 04:18 am (utc+8)
The artistic style has been ft base the Pony XL model, following the guidelines of the Animagine XL V3 caption rules, by training with the collective works of 55 artists. A detailed inventory is included, although not every term may yield effective results.
Sorry, a mistake yesterday,in the final step I merge 5 layers from js2pony. I tested it and the difference isn't significant, so I don't feel like re-uploading.
Let's just go with the flow (●'◡'●).
please add artist name in the list to prompts if you want to get some style.
use prompts like use pony XL
artist list in dataset(sort by random?):
hoji (hooooooooji1029)
fuzichoco
kani biimu
kfr
mika pikazo
scottie (phantom2)
toosaka asagi
atdan
dsmile
gin00
happoubi jin
hiten (hitenkei)
kazutake hazano
ke-ta
kedama milk
mignon
misaka 12003-gou
momoko (momopoco)
morikura en
noco (adamas)
shiratama (shiratamaco)
torino aqua
amashiro natsuki
anmi
as109
ask (askzy)
fangxiang cuoluan
kaede (yumesaki kaede)
matanonki
mochizuki kei
rin yuu
soraneko hino
tiv
toraishi 666
toridamono
ciloranko
ogipote
poco (asahi age)
jima
miv4t
noyu (noyu23386566)
quan (kurisu tina)
rurudo
sy4
hito komoru
snow is
weri
akakura
binggong asylum
chen bin
nekojira
akizero1510
hong bai
da mao banlangen
quality_tag_list(in this ft caption):
"masterpiece",
"best quality",
"great quality",
"good quality",
"normal quality",
"low quality",
"worst quality"
year_tag_lsit(in this ft caption):
2005 <= year <= 2010: year_tag = "old"
elif year <= 2014: year_tag = "early"
elif year <= 2017:year_tag = "mid"
elif year <= 2020: year_tag = "recent"
elif year <= 2024:year_tag = "newest"
dataset from:
https://huggingface.co/datasets/KBlueLeaf/danbooru2023-webp-2Mpixel
https://huggingface.co/datasets/NebulaeWis/gelbooru_images
images process and filter : https://github.com/deepghs/waifuc/
caption & creation dataset : https://github.com/KohakuBlueleaf/HakuBooru
other code :me
License
This model is released under Fair-AI-Public-License-1.0-SD Plz check this website for more information: Freedom of Development (freedevproject.org)