Illustrious model, for now tested in hassaku XL, since no one was doing an illustrious of Gu Yuena i decided to do it myself....
Description
FAQ
Comments (19)
大佬来点姜青娥 长公主和采薇的谢谢
Jiang Qing'e is on the agenda, i can't promise when but i plan to finish it after i get the time
Hello!I am a starter of lora trainimg.I want to know how you use only two trigger words to make a lora about a particular character with specific costume.I usually spend a long time tagging during training the model,and even the lora I make is not very good. If possible,could you leave a contact information to guide me on training lora?I will appreciate it if you could help me.
so the trick is this, it's simple, while you caption the images, let's take guyuena as an example, the most important thing is to add the keyword "cropped" in all her images,
next
i don't add keywords to describe her such as white hair, silver hair, purple eyes, very long hair, don't add them only " guyuena, 1girl, " is sufficient, only focus on what is different in other images such as sitting, standing, from_behind, from front, upper body, full body, etc, if background is different add it too.
about the outfit, i just add caption of "default_outfit" in all the images, i didn't add any other keyword such as "dress", "white dress", "skirt", "thighhighs" etc. If you add those it will not focus on the "default_outfit", that's why my Su Yan lora is inconsistent with the outfit also don't add "default outfit", the "_" is important.
try these first and see how it goes and let me know how it goes.
Also this is how i captioned one of my images, the others follows a similar pattern, it depends on the image-
"guyuena, 1girl, closed mouth, default_outfit, upper body, one eye closed, window, door, cropped"
@SHADOW_D_X How many training datasets are suitable?
@Ruinor the more the better, but if you are experimenting first i would suggest trying with 50-70 good enough images, make sure to focus only on the character you want to make, my first lora Ha-young was 156 images, but some images had lower pixels and overall as a result a bit overtrained. Su Yan and Xue Yue lora was 71-75 images i cropped out and screenshot from pc the specific characters also zoom in as much as possible while taking screenshot, it turned out decent. Gu Yuena was 41 images, but all the images were extremely high quality that i got from other AI generations of Gu Yuena that i liked and combined them.
@SHADOW_D_XI see.Thank you very much.
@Ruinor No worries, if you run into issue let me know
可以训练从姑获鸟开始中的丹娘嘛(´･_･`)
???
@SHADOW_D_X I am very sorry, I opened the translation software, and I would like to ask you if you can make the Danniang in the cartoon "All starts with ubume".The Chinese name of the manga is 《从姑获鸟开始》. If you do not have time or are not interested, please consider me as disturbing you.
@asdqrwe sorry i don't have the intention to make the character now, i will look into the series first.
@asdqrwe 你为什么不直接打搜索一下吗？我早就上传过了
@god2356 and illustrious version too wow
@god2356 我询问是14天前，而你上传是12天前，那时你还没有上传
@asdqrwe yeah the time period is very short, and searching everyday for a specific character is a pain
@SHADOW_D_X Yes, even your favorite characters can't be checked every day