中文版在后面。
English Version
For convenience, set N=repeats * count. 6_characterA has 20 img inside so the N is 120.
You know, the more tags you delete, the fewer "trigger words" you might need to memorize and use. But it may be also harder to change clothes if you do not add other clothing's data, and to use it with other clothing loras.
Well for those automatic bots…… Web Crawler, Start! Some newbies also. Throw all those data collected by crawler into the pot and start cooking. Then you might find lots of "trigger words" and the character's features somewhat mix up.
PS. someone even inputs a long prompt in the "trigger words", with his/her name.
If we want a character lora that can not only get her official costume but also put on other costume, there are different ways to make a character lora easier to change clothes:
Keep clothing tags. This character does not have specific clothes, or you just want to use it for changing clothes( like the madoka lora I use) , or his/her clothes is very simple,
or he/she do not wear anything like Quinella from SAO.So just do not delete clothing tags in datasets."Portrait Folder". This method was told by L_A_X(https://www.bilibili.com/video/BV11h4y1x7LH). One folder for the character you want to train, and delete nearly all the feature-related tags about him/her like xx eyes, xx hair and clothing, put the main trigger word at the very front. The other folder for the portraits of the character (yeah, you might have to crop images to get those portrait), just delete tags about his/her head's features, and use another trigger word. Make N of normal folder and N of portrait folder as close as possible.
However this might not be perfect somtimes…For example this Platinum (Swimsuit) image I input latex bodysuit with lora weight 1, but as you can see there is not latex bodysuit.
And I am lazy I do not want to crop lots of images
"Other Folder". One folder for standard outfit and the other folder for different clothing. Use different trigger words and balance their N of course. You might have use this kind of loras like Bloodsuga's Taihou(Azur Lane), zbw's Arknights loras.
However, not all characters have various skins or are popular enough to have data for other clothing.
"1 Repeat Manual Regulation".
Take Closure (Arknights) for example. Originally I though her clothes are not very complex, "shirt, open jacket, belt" mainly.
But……China dress one, weird. Latex bodysuit one, where is it?
Do not expect that AI can always "think" as what you think or tagger "thinks"(
Though Closure has 280+ images on danbooru, most of them describe standard outfit, after all she do not have official alternative costume(
N of Closure folders is about 150. Then I added 34 images of different girls wearing different clothes like long dress, one-piece swimsuit, kimono, bikini, birthday suit, giving them 1 repeat.
As you can see, the ability of changing clothes is improved a lot.
Of course you can combine these methods together. "completely nude" can also be regarded as a "special costume".
中文版
方便起见 设 N=重复 * 图片数。比如6_characterA这个文件夹有20张图，那么就是120。
众所周知，删的tag越多，要记和用的触发词可能就越少。但也可能会你在不添加其他服装数据的情况下更难换衣服和搭配其他服装lora。
不过对于那些自动炼丹机器人……网络爬虫，启动！还有一些萌新也是。把爬虫收集来的数据一股脑地丢进锅里炖了。结果你可能看到有一大堆“触发词”，并且角色特征也有点混。
PS. 有的家伙甚至在页面的触发词那一栏塞了长长的prompt，甚至还带着自己的名字。
如果我们想要一个能还原原设服装的同时还能换装的角色lora，以下是提高换装能力的一些方法：
服装全标啥也不删。这个角色没有特定的衣服, 或者你本就是用来换衣服的只练头，或者ta的衣服特别简单
，或者跟SAO的葵涅拉一样压根不穿衣服。这种情况就没必要删衣服标。"头部文件夹"。五月老师教的(https://www.bilibili.com/video/BV11h4y1x7LH)。一个文件夹装角色，服装标全删，把触发词放最前边。另一个文件夹放角色的头像（当然一般你得裁出那么多头像来），只删头部特征标，用另一个触发词。平衡两个文件夹的N。
然而有时这个法子并不完美，比如这张泳装白金我在权重1时输入latex bodysuit你看根本就没有乳胶衣。
而且我懒得裁一堆图。"其他服装文件夹"。一个文件夹装标准的，一个文件夹装不同衣服的。用不同的触发词并且显然平衡一下N。你可能用过这种lora，比如圣姨哥的大凤，zbw（空城大佬）的粥lora。
然而不是所有角色都有一堆皮肤或是很热门有其他衣服的同人图。
"垃圾（手动）正则法".
拿可露希尔举例。本来我觉得她的衣服挺简单的主要就"shirt, open jacket, belt"。
但是你也看到了，旗袍怪怪的，乳胶衣也不见踪影。
不要期望AI想的能跟你或者打标器一样啊喂(
虽然奸商在danbooru有280多张图但是大多数都是原皮，
毕竟本来她也没其他皮肤（啥时候入池可露希尔文件夹的N是150+。然后我加了34张其他妹子的图，她们穿着不同的衣服，如长裙，连体泳衣，和服，比基尼，还有啥也不穿，并给了1重复。
换装的能力这就明显改善了。
当然你可以把这些法子结合起来。"completely nude" 也可以被视为特殊的服装。