Nakatsukasa Kirino/中務キリノ (Blue Archive) for PonyXL (3 Outfits)

————v2.0beta————

v2.0beta 主要用于生成“泳装”和“偶像”（在第5 PV中出现）两套服装，生成“警服”请仍使用v1.1。

v2.0beta 主要用于生成“swimsuit”和“idol”（出现在5th PV中）的服装。生成“police uniform”请继续使用v1.1。

由于偶像服的数据集过少，未能成功还原发型，但除此之外的效果都出乎意料地好。

Due to the limited dataset for idol outfits, I was unable to successfully replicate the hairstyles, but apart from that, the results were unexpectedly good.

遗憾的是，官方服装反而过拟合了，如果需要生成官方服装我强烈推荐使用v1.1。如果要使用v2.0beta生成官方服装，请将CFG Scale降低至4.5，并配合lora block weight使用：lora:NakatsukasaKirinoXL:0.8:0.8:lbw=1,1,1,1,1,1,1,1,1,0,0,0。

Unfortunately, the official costume "police uniform" seems to be overcooked. If you need to generate official costume, I strongly recommend using v1.1. If you still want to use v2.0beta to generate official clothing, please decrease the CFG Scale to 4.5 and use it with LoRa block weight extension: lora:NakatsukasaKirinoXL:0.8:0.8:lbw=1,1,1,1,1,1,1,1,1,0,0,0。

——

泳装触发词（Swimsuit trigger words:）：

kirino \(swimsuit\) \(blue archive\),

sandals, police hat, swimsuit, braid, blue bikini, side-tie bikini bottom, navel, twin braids, blue vest, blue headwear, open clothes, crop top, shirt... megaphone, whistle...

——

偶像服是一套固定服装，只需要一个触发词（The idol outfit is a fixed set of clothing, requiring only one trigger word）：kirino \(idol\) \(blue archive\)。

增加以下tag会使结果更加稳定（Adding the following tags will make the results more stable）：pink puffy short sleeves, pink bow, mismatched legwear, asymmetrical legwear, striped pink footwear, pink shoes。

——

注：我试图重新调整repeat值来避免官方服装过拟合的现象，却影响了其他两套服装的效果。因此，我所提供的训练集并非v2.0beta使用的，而是经过了repeat值和极少量的图片的调整。未来“桐乃（偶像服）”实装后，我会继续更新这个lora。

Note: I attempted to readjust the repeat value to avoid overcooking of the official costume, but it affected the effectiveness of the other two sets of outfits. Therefore, the training set I provided is not the one used by v2.0beta, but rather has undergone adjustments to the repeat value and a minimal amount of images. After the implementation of "Kirino (Idol)" in the future, I will continue to update this LoRa.

————v1.1————

更小，更灵活，也许效果更好

Smaller, more flexible, perhaps yielding better results.

请视情况（和视角）选择触发词：

Please choose trigger words based on the context (and perspective).

触发词 (Trigger word)：

kirino \(blue archive\)

头部触发词 (Head trigger words)：

braid, twin braids, long hair, hairclip, hair between eyes, bangs, ahoge, hat, multicolored hair

上半身触发词 (Upper body trigger words)：

white gloves, blue necktie, long sleeves, armband, white jacket, badge, holster, belt, cylindrical objects attached to belt, belt pouch

下半身触发词 (Lower body trigger words)：

white boots, skirt, black pantyhose

我想知道怎么在同一个角色lora中放入两套服装。如果您愿意分享相关教程，我会非常感激！

I want to know how to put two costumes of the same character in one LoRa. I would greatly appreciate it if you could share any relevant tutorials!

————v1.0————

这个lora只能基于pony及由其衍生的底模生效。

触发词：kirino \(blue archive\)

可选的触发词（我想不止这些）：hat, gloves, necktie, gloves, necktie, long sleeves, armband, jacket, badge, holster, pantyhose, skirt, belt, cylindrical objects attached to belt, boots, multicolored hair, handgun, holding weapon

推荐权重：0.8。

推荐的CFG scale：我也不知道！使用pony v6底模时，我设置为6.5或7。但在使用autismix底模时，我不得不下调至4.5，否则画面线条会非常粗。

推荐使用lora block weight插件：lora:NakatsukasaKirinoXL-000029:0.8:0.8:lbw=1,1,1,1,1,1,1,1,1,0,0,0

这是我第一次训练lora，可以预想到会有许多不足之处。事实上我甚至无法判断它究竟是否有欠拟合或过拟合的情况。

我分享了训练数据。如果您愿意教我如何优化训练数据，我将感激不尽。

----AI Translation----

This LoRa only works based on pony and its derivatives.

Trigger word:kirino \(blue archive\)

Optional trigger words (I think there are more than these): hat, gloves, necktie, gloves, necktie, long sleeves, armband, jacket, badge, holster, pantyhose, skirt, belt, cylindrical objects attached to belt, boots, multicolored hair, handgun, holding weapon

Recommended weight: 0.8.

Recommended CFG scale: I don't know either! When using the pony v6 base model, I set it to 6.5 or 7. However, when using autismix, I had to lower it to 4.5, otherwise the lines in the image would be very thick.

Recommended to use the LoRa block weight plugin: lora:NakatsukasaKirinoXL-000029:0.8:0.8:lbw=1,1,1,1,1,1,1,1,1,0,0,0

This is my first time training a LoRa, and I can foresee there will be many shortcomings. In fact, I can't even determine whether it is underfitting or overfitting.

I have shared the training data. I would greatly appreciate it if you could teach me how to optimize the training data.