Lora that mimic the style of Itou
Trigger word is "9itou_onsoku " for the lora
This is a reinforcement LoRa, the base model (Anima base V1) has already a good knowledge of the artist.
But the artist style as evolved and the base model seem of bias into it's older style (2018 2012) even with new and newest tags. It also does no draw like the artist does the android seams on skin too.
Older style has lot of roundness on the character and are often highly illustrated <- base model
Newer style has more edge and more anime looking. <-LoRa
Example : left is no trigger , middle is @itou \(onsoku tassha\) , right is the lora 9itou_onsoku
There is also a compare picture in the showcase were you can see the difference how base model create android skin vs the one with the lora (the girl straddling a boy)
Long story short and random information about prompting:
you can call upon base model knowledge with “@itou \(onsoku tassha\)”
you can call upon lora with recent style and characters knowledge with “”9itou_onsoku " or "9itou_onsoku score_7_up"
You can stack both but don't expect sure fire result, you will need to do weight tuning
The Lora is better at giving the recent style from the artist than the base model
Trained locally with my own tool, ComfyUI-AnimaTrainer
Description
First version was subpar and was not published
dataset 2 with mixed style and repeat x2 on newer assets
"Deep" learning - Last Epoch, run with regularization, Do inference between 0.72 and 0.86 weight to keep good prompt adherence
in ComfyUI use "model sampling aura flow" before Ksampler with high value 8 to 12 for complex compositions !