Starry XL enhances the Kohaku Epsilon model by specifically targeting the styles of Pixiv's top artists and expanding the character dataset to generate high-quality waifu artistic imagery.

The model has undergone extensive data balancing and processing, ensuring effective representation of approximately 600 of the most popular artist styles(total num of artists is far more than 600). Additionally, it has received enhanced training specifically targeting characters that the Kohaku Epsilon model performed poorly on in CCIP evaluations(approximately 1200 character with low ccip scroe ).

After a series of processing steps, the final training set consists of 626,495,use rtx 4090 for training 6epochs over 450 hours.

Model Information

• Developed by:  me

• Funded by: me

• Model type: sdxl

• Finetuned from model:Kohaku-XL Epsilon

• License: Fair AI Public License 1.0-SD

Usage Guide

Starry is based on epsilon, and during training, the caption are overall close to Kohaku epsilon, so the overall usage is the same

There is a wildcard for 600 artists in the attachment.

starry_aritst_600_list

for other artists and characters, please use the existing list from Kohaku Epsilon. https://civarchive.com/api/download/models/445973?type=Training%20Data

Note that Starry requires high accuracy in artist names, so ensure there are no spelling errors and use the correct artist/character tags.

<1girl/1boy/1other/...>, 

<character>, <series>, <artists>, 

<general tags>,

<quality tags>, <year tags>, <meta tags>, <rating tags>

Quality tags: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality

Rating tags: safe, sensitive, nsfw, explicit

Date tags: newest, recent, mid, early, old

Recommended Negative Prompts

If special styles are required, need to modify Negative Prompts

long:

bad anatomy,blurry,(worst quality:1.8),low quality,hands bad,face bad,(normal quality:1.3),bad hands,mutated hands and fingers,extra legs,extra arms,duplicate,cropped,text,jpeg,artifacts,signature,watermark,username,blurry,artist name,trademark,title,multiple view,Reference sheet,long body,multiple breasts,mutated,bad anatomy,disfigured,bad proportions,duplicate,bad feet,artist name,ugly,text font ui,missing limb,monochrome,  

short:

nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name,

Style Select

you can directly use artist's prompt to generate image.

1girl,momoi \(blue archive\), blue archive,

{style},

solo, headphones, halo, pink halo, white jacket, short hair, bow, shirt, necktie, white background, white shirt, blue necktie, fake animal ears, animal ears, pink bow, collared shirt, simple background, pink eyes, blonde hair, animal ear headphones, looking at viewer, hair bow, jacket,newest, masterpiece, best quality,  absurdres, highres, 

You can also used DanTagGen to generate images with a strong style from an artist

1girl,{style}, {dtg expand} newest, masterpiece, best quality,  absurdres, highres,   

hito komoru

aris \(blue archive\), blue archive,quan \(kurisu tina\), portrait,zoom layer,

kani biimu

rurudo

Combining multiple artists is highly recommended, and you can use the artist list to try different orders and combinations. In fact, you can use the famous nai3 artist prompts to combine styles directly. (This is not a simple nai3 distillation, it uses artist prompts for style combine)

(ningen mame:0.9), ciloranko, sho \(sho lwlw\), (tianliang duohe fangdongye:0.8), ask \(askzy\), wlop, 

Acknowledgements

https://civarchive.com/user/narugo1992 Thank you, Jerry, for sharing all the tools, datasets, and other resources. This includes image processing tools and related databases. Without the resources provided by Jerry, the training work would have been much more challenging to carry out..

https://civarchive.com/user/kblueleaf  Special thanks to KohakuBlueleaf for developing and providing the Lycoris and Hakubooru datasets, and most importantly, for training the remarkable Kohaku Epsilon model.

https://huggingface.co/cagliostrolab I would like to express my gratitude to cagliostrolab for sharing the insightful Tag Ordering Rules, which have proven to be incredibly helpful for training purposes.

https://github.com/kohya-ss Training scripts

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)