新人不懂事练着玩的，比较接近原作者的效果，我感觉很不错，也可以不加提示词拉低权重当光影lora使用。个人推荐权重: 0.65-1.0，底模我推荐RDBT 0.39b 或者 RDBT v2，对于大白兔的风格提升度很高，对其他的画风lora适配性也很高！它可以有效的改善这个lora的朦胧感和光影。

但是接近也代表着过拟合，已知问题有：

1. 格调可能会出现轻微薄码(不是必出)，如果有出现马赛克请拉高 uncensored 权重 (uncensored:2) 可以解决。

2. 虽然我有打标床、枕头、窗帘等，但是某些场景下需要你更详细的写出周围的环境物品，例如你写一个角色躺下 (lying) 的姿势，背景是 city, street 之类的，但是却没有描写躺在地上(lying on ground)，那么出图效果会变成大街上凭空多了一张床，而角色躺在床上。

3. 一些角色罚站（bushi）场景可能不加 full body 就不会出全身照，如果加了 full body 效果也不佳，同上 (full body:2)

4. 出双人角色对提示词要求更严格，你需要分开对每个角色进行特征描写，直接混在一起出图会串特征。

5. 碧蓝档案的普拉娜或者阿罗娜有一定概率出爱心光环，不想出请将 heart halo 加入负面，（heart halo:2）可以解决这个问题。(当然你不出这个角色可以忽略这个问题）。

6. 把光美露露卡的部分特征给学了进来，可以通过预览图得到提示词。

7. 如果觉得画面亮，是因为拟合了素材图像里面的光影，原图就很亮这是无法避免的，所以生图的时候总有一个角度会有光照到角色身上，可以使用一些增加暗度的lora来对冲。

（碎碎念：我试图将鸣潮姬头四那张图打上标去实现效果，但似乎效果不是很行，遂放弃没有加进来，但是原神三罗那张图是有写的，只不过发不出来就是，你们可以尝试反推去出。）

Trigger Words：@dabaitu

A newbie just messing around with training this for fun. It stays relatively close to the artist's output, which I personally think works quite well, also you can also use it at a lower weight without any trigger words, essentially as a lighting/shading LoRA. weight: 0.65–1.0. For the base model, I recommend RDBT 0.39B or RDBT V2. They provide a significant boost to this LoRA while also being highly compatible with other style LoRAs. It can effectively reduce the hazy look of this LoRA and improve its lighting.

However, staying close to the artist also means that it is somewhat overfit. Known issues:

1. The penus may occasionally show slight mosaic/pixelation artifacts(It does not necessarily produce mosaics). If you encounter noticeable mosaic artifacts, try increasing the uncensored weight to (uncensored:2), which can resolve the issue.

2. Although I tagged things such as bed, pillow, curtain, etc., in some scenes you may need to describe the surrounding environment and objects in more detail. For example, if you prompt a character with the pose lying and use a background such as city, street, but don't specify that they are lying on ground, the generated image may end up placing a bed in the middle of the street, with the character lying on it.

3. In some standing pose scenes, the model may not generate a full-body shot unless you include full body. However, adding full body does not always produce ideal results either. Similarly, you can try increasing its weight to (full body:2).

4. Generating two characters requires more precise prompting. You should describe the characteristics of each character separately. If you simply mix all the character descriptions together, their features may bleed into each other.

5. Plana and Arona from Blue Archive has a certain chance of generating a heart halo. If you don't wanna it, adding heart halo to the negative prompt,（heart halo:2）can fix this issue. (Of course, if you aren't trying this character, you can simply ignore this issue.)

6. It learned Pretty Cure Ruruka's characteristic features from image. You can get the prompts from the preview images.

7. If you feel that the image is too bright, it’s because the LoRA has learned the lighting from the training images. The original images are already very bright, so this is unavoidable to some extent. As a result, when generating images, there will often be an angle where light falls directly onto the character. You can use a LoRA that darkens the image to counteract this effect.