The Latest version is: t007.00e01
The LoRA was created to be used with NAI-Based "2D / Anime models" at 0.4-0.6 weight to create beautiful, curvy 2D/2.5D girls.
Trained on top of NAI (nai-animefull-final-pruned). Low weight is needed not because the lora is overcooked or overfitted, but because you'll get realistic results otherwise...(so dont dial up, unless it's your goal. This LoRA will make any 2D scene into 2.5/3D if turned high up enough)
2042 pics were used for the training
Dataset/Tags was manually curated
LoRA (ver. t001.00e05) was trained for a crazy 36075 steps ("36075 img x repeats" / "5 batch size" * "5 epochs") and then fine tuned to 60k+ steps (ver. t007.00e01) on DAdaptation. The whole thing was more an excessive experiment of mine where I tried different things in 4 different trainings.
This version is a resized 128dim/128alpha version of a 256dim/256alpha LoRa
Few of the manually tagged things:
"precise tanlines" for signature tan, 80s haircut, 90s haircut, updo haircut, casual haircut, short messy hair, areolae slip,"shiny, shiny skin": for oily look, etc.
NOT INTENDED FOR REALISTIC MODELS OR BASE 1.4/1.5 models.
NOT FOR COMMERCIAL USE
DO NOT USE IT FOR ANYTHING ILLEGAL
Description
Easter themed release ^^
3 additional trainings on top of t001.00e05 LoRA.
Final version as far as i know. The whole thing was more an excessive experiment than anything else.
The last 2 epochs were trained at different resolutions.
Ver t006.00 was trained only on landscape-AR images at 1024x672
Ver t007.00 was trained only on portrait-AR images at 672x1024
This is 128dim resized/LITE version of the original 256dim LoRA.
Switched to Anything3 for training midway as a test, but tracked back to NAI.
(It looked very good on some models, but compatibility to "ANIME Models" got much worse after the switch. So went back training on NAI)