准备了挺长时间制作的一个lora，但是很多地方的效果仍不尽人意，比如特定怪物的造型，打标的准确性，黑色的边框，泛化性不佳等等。以后有时间再优化吧

基本的提示词框架：

sinisistar2 style, split panel layout, (表示sinisistar2画风+左右构图)

left part of image:character portrait panel, 1girl, lelia, ...(左构图+角色插画画风+具体内容)

right part of image:pixel gameplay panel, pixel art, 1girl, lelia, ...(右构图+像素游戏画风+具体内容)

你可以使用纵向构图+提示词：sinisistar2 cg style, character portrait panel, 1girl, lelia,

的方式来单独画立绘

也可以：sinisistar2 pixel style, pixel gameplay panel, pixel art, 1girl, lelia,...

的格式来单独画像素横版卷轴部分

English ver：

I spent quite a while creating this LoRA, but the results are still unsatisfactory in several areas—such as the designs of specific monsters, tagging accuracy, the presence of black borders, and poor generalization. I’ll look into optimizing it further when I have the time.

Basic prompt framework:

sinisistar2 style, split panel layout, (indicates sinisistar2 art style + split-screen composition)

left part of image: character portrait panel, 1girl, lelia, ... (left-side composition + character illustration style + specific content)

right part of image: pixel gameplay panel, pixel art, 1girl, lelia, ... (right-side composition + pixel game art style + specific content)

You can generate standalone character portraits using a vertical layout with these prompts: sinisistar2 cg style, character portrait panel, 1girl, lelia,

Or generate standalone pixel-style side-scrolling sections using this format: sinisistar2 pixel style, pixel gameplay panel, pixel art, 1girl, lelia, ...

目前最大的问题：泛化性差

1.人物不可替换性：使用别的人物触发词如mika，会带有大部分部分lelia的装饰和特征

2.由于素材全都是cowboy shot，导致模型无法画出腿部

3.像素部分的打标描述准确性不够，需要完全的人工审核，外加画面内容不易于模型理解，导致这部分的内容画的比较抽象，有时间再做吧