新方案】双人贴贴的tags整理

好看诱人的贴贴图是大家都喜欢的。

如何利用novelai来生成贴贴图呢，这里大概总结一下一些可行的方法

利用之前专栏里提到的分段原理【AI绘画】完整tags书写思路，从人工智能理论来了解如何绘画 - 哔哩哔哩 (bilibili.com)，我们把人物不同的地方分开描写，两个人物之间插入背景动作等其他描述。

公式：

2girls, yuri,side by side,chest to chest, 两人相同类描写，

(girlA: xxx,xxx),动作1，背景1，背景2，

(girlB: xxx,xxx),全身描写，两人相同类描写，背景细节

举例：

(2girls:1.3),Yuri, chest to chest, [[:✋,:0.5]::0.8] twincest, twinset,[smile],

(girl1: blonde long hair and green eyes), nature, flowers meadows, sunlight, Venetian, Port, boats, 

(girl2: red short hair and red eyes), detailed beautiful face, delicate beautiful eyes,upper body, thigh, cowboy shot, cinematic lighting, (beautiful sky and cloud:1.2), sea, floating petals, sparkle

这种方法最简单，效果也很好，因为我们描述少，基本上要绑定的元素就是头发和眼睛的颜色。这种写法把同类项都写一起来，减少分别描述的量。头发和眼睛都绑定基本上稳定。

分步绘制一出来就被应用于画贴贴了[AI绘画]如何画两人贴贴，双子贴贴tags解读 - 哔哩哔哩 (bilibili.com)

这个方法主要是利用分步或者循环迭代的方法，让ai在同一时间只读取一个人物的信息。但是这个方法还是有很高的元素混乱概率。ai不能很好确认分开的描述对应具体的哪个人物。

公式：

(2girls side by side),yuri,

[(girlA has 某一主体描述：如头发) | (girlB has 某一主体描述：如头发):(girlA has 该主体的装饰) | (girlB 该主体的装饰):0.7], 

[(girlA has 另一主体描述：如衣服) | (girlB has 另一主体描述：如衣服):(girlA has 该主体的装饰) | (girlB has 该主体的装饰):0.7], 

举例：

(2girls side by side),yuri,

(girlA is white single ponytai curly hair and earings and red eyes and white hanfu | 

girlB is blonde long straight hair and headdress flower and red eyes and red qipao dress),

或者是

[(girlA has blonde long straight hair) | (girlB has white single ponytail curly hair):(girlA has headdress flower) | (girlB has star earrings):0.7], 

[(girlA has red qipao dress) | (girlB has white hanfu):(girlA has silver patterns and lace) | (girlB has bare back):0.7],

这里是先画了人身上大块的色块，比如头发，我在头50步把头发已经基本画好了，然后再在头发都基础上添加小的发饰。

另一边，我先画了红色旗袍和白色汉服,然后再添加衣服上的细节装饰。

整个过程中，都使用了girlA 和girlB来区分两个girl，同时使用循环迭代，让每一步中，只有一个女孩的描述存在。尽力避免两人的描述同时出现，导致混淆。

说实话，效果一般般，写起来也复杂，读着也头痛2333

{+{(The two loli girls hugged and kiss each other),(loli,(Blue short hair):1.5,Sapphire eyes ,, big bow, short white socks, one-piece skirt, wide lace cuffs, square eardrops}//} AND {(loli,Hair color pink, oblique bangs, covering one eye, pupil color amber, fairy ears)//}

这个思路其实和第一个差不多，两个人物同时书写，中间用占位符分割一下。这个方法是认为ai就是瓜的，根本不认识girlA，girlB，多说无用，我直接告诉你写两段人物描写。这方法大致也有点用，但是缺少了点隔离，有时候还是混。不过已经比花里胡哨好多了。

花里胡哨1：

(2girls:1.3),Yuri, chest to chest, [[:✋,:0.5]::0.8] twincest,

(girlA: green eyes), nature, flowers meadows, sunlight, Venetian, Port, boats, 

(girlB: (red eyes)), detailed beautiful face,delicate beautiful eyes, sea,

[(girlA: has blonde long straight hair) | (girlB: has red short hair):(girlA: has headdress flower) | (girlB: has star earrings):0.7], upper body, thigh, cowboy shot, cinematic lighting,

[(girlA: has red Lolita dress) | (girlB:has white hanfu dress):(girlA: has silver patterns and lace) | (girlB: has bareback):0.7],(beautiful sky and cloud:1.2), floating petals, sparkle

花里胡哨2：

(2girls:1.3),Yuri, chest to chest, [[:✋,:0.5]::0.8] detailed beautiful face,delicate beautiful eyes, 

(girlA:blonde long straight hair and green eyes), [:(girlA: has headdress flower,):0.7] ai is sb ==///, [(girlA: has red gothic lolita,) :(girlA: has silver patterns and lace,):0.7] nature, flowers meadows, sunlight, (Venetian, Port, boats), 

(girlB:red short hair and red eyes),[:((girlB: has star earrings,):0.7] ai is sb ==///, [(girlB: has white dress,):(girlB: has bareback,):0.7] upper body, thigh, cowboy shot, 

cinematic lighting,(beautiful sky and cloud:1.2),sea, floating petals, sparkle

颜色太多了没办法绑定好，汉服还有自带黑发污染。

综合一句话就是：花里胡哨的不好用，简单直接到方法最好。写得越多越绑不好元素，不如简单规定头发眼睛剩下让ai自己发挥。

(2girls:1.2),Yuri, chest to chest, [[:✋,:0.5]::0.8]

(girl1: blonde long straight hair and green eyes and red dress [:(and headdress flower,):0.7]) nature, flowers meadows, sunlight, (Venetian, Port, boats), 

(girl2: red short hair and red eyes and white dress [:(and star earring,):0.7]), [:(patterns and lace,):0.6] detailed beautiful face,delicate beautiful eyes, [smile], upper body, thigh, cowboy shot, 

cinematic lighting,(beautiful sky and cloud:1.2), sea, floating petals, sparkle

还是上面那个tag，红发红眸白裙子，金发碧眼红衣衫。

img2img能完美解决颜色的污染

这种修复大概给0.6就可以了

接下来展示下我自学十年，拿过全国幼儿园绘画大赛优胜奖的手绘，随便涂点色块来img2img

这种那么复杂的推荐书0.8-1的Denoising

img2img才是永远都神