v2.0
Trained by 60k images from Danbooru and Gelbooru + 8k AI-generated images.
・UNet lr= 5e-6 ~ 3e-6
・Text Encoder 1 Lr = 3e-6 ~ 1e-6
・Text Encoder 2 Lr = 2e-6 ~ 7e~8
To prevent excessive training, the dataset categorized by style sets repetition numbers based on the amount of data.
Prompting Guide
・1girl/1boy/2girls..., characters, copyright, style, general tags, rating, score_9, score_8_up, score_7_up.
Negative Prompts
・score_4, score_5, score_6, source_pony, source_furry, monochrome, realistic, rough sketch, fewer digits, extra digits
Score tags are adjusted so that, average anime illustrations are generated, without the aesthetically tuned styles like in v1.0.
To increase detailing or change styles, you can use prompting similar to 'NovelAI' or use 'lora'.
Some artist tags seem to work. When using artist tags, write the score tags at the end.
Example
v1.0
Introduction
This model is finetuned from Pony Diffusion V6 XL.
Trained by SFW and NSFW images scraped from Danbooru, Gelbooru.
Tagging follows the template of Pony Diffusion.
score, 1girl/1boy/2girls..., characters, copyright, style, general tags, rating
Negative Prompts
score_4, score_5, score_6, source_pony, source_furry, monochrome, 3d, photo, hyperrealistic, realistic, rough sketch, fewer digits, extra digits, signature, artist name
Score tags have been made clearer.
score_9: high quality illustrations
score_8_up: realistic texture
score_7_up: anime-like
Although no training focused on specific characters, some characters may have been learned.
License
Support☕ https://ko-fi.com/sfa837348