replicate the character design in all the style and aspect you want.
have been train on both illustration + 3D model provided by commisionner Tomy117
dataset of 335 VRC screencap + 30 illustrations from various artists including some of my own gens.

due to lack of dataset regularization on object interaction, the lora perform less well when we try to occlude the body.

this model works the best when prompted "blue demon, eydis woman" using a merged model of Noobai (not really illustrious) such as my favorite KonpaEvo MIX or the popular NTR mix.

experimental parameters for DORA (LYCORIS) using Prodigy optimizer instead of my still favorite ADAMW.
3H train on 3090 for 1570 steps of 5 batches. 85% power draw

provided training parameter as JSON file