Murasaki Shion (Hololive) LoRA

Thank you for deleting my previous LoRA without any warning (I discovered it delection 1 month later) because "Hololive" and "gothic_l*lita" contain a banned word and excluding the domain from wayback archive means that I can't copy paste my old LoRA description (this is annoying for recommended prompts)

01-05-2023/V2 Update:

• Reduce the model file size to 9MB.

• Use a way lower network dim (8 instead of 128) and Alpha (1 instead of 128) to prevent overfitting

• More flexible

• Rebalanced dataset (I uploaded it)

• More accurate official swimsuit

• Add a new token for each canon outfit (using booru token was a bad idea):
  ShionDefault

  ShionSwimsuit

  ShionGothic

V1 is a reupload of my prevous LoRA

This is a LoRA, this mean that you need to update Voldy/AUTOMATIC1111's webui to use it (use git pull) .

Weight:

You won't need to use a lower weight unless you want to use multiple LoRAs

Recommended prompt:

That LoRA can reproduice some of her models

For V2:

Default outfit:

black gloves,single hair bun, witch hat, vertical-striped_shirt,pinstripe_shirt,midriff, crop top,miniskirt, ShionDefault

Gothic L*lita :

twintails, animal_ear_fluff,gothic_lolita,lolita_fashion,halter dress, cat ears,cat_girl,long dress, halterneck, layered dress,frilled dress, center_frills,detached sleeves, ShionGothic

Official Swimsuit :

bikini, bikini_skirt, black_bikini,criss-cross halter, swimsuit, ShionSwimsuit

For V1:

Same thing as V2 but without ShionDefault,ShionSwimsuit and ShionGothic

Model used for the example:

V1 examples : AOM 2 and Pastel Mix (model version) .

V2 examples : AnythingV4.5

描述:

247 images

6 epochs

4188 steps (batch size:1)

scheduler: cosine

Text encoder learning rate: 1e-5 

Unet learning rate: 1e-3 

max bucket resolution: 512 

clip skip: 2 

dim: 8

alpha: 1 

batch size=1 

Base model: NAI

训练词语: ShionDefault,ShionSwimsuit,ShionGothic