AI绘画Stable Diffusion技巧分享

Prompt运用规则及技巧

AI绘画Stable Diffusion技巧分享资源：

1. https://publicprompts.art/（最适用于OpenArt 线上模型 https://openart.ai/）

2. https://docs.qq.com/doc/DWFdSTHJtQWRzYk9k（转自元素法典，含各类风格用词）

3. https://www.ptsearch.info/home/（实例查询）

4. http://prompttool.com/NovelAI（体验还不错）

使用：

1. 通用Tag起手式：

正向：

masterpiece, best quality, 更多画质词，画面描述

反向：

nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers,extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,根据画面产出加不想出现的画面。nsfw（not safe for work＝涩图）

2. 越靠前的Tag权重越大；比如景色Tag在前，人物就会小，相反的人物会变大或半身。

3. 生成图片的大小会影响Prompt的效果，图片越大需要的Prompt越多，不然Prompt会相互污染。

4. 在Stable Diffusion 中使用()英文括号可增加括号中Tag在画面中的权重 x1.1，[]可减小Tag权重x0.91。在NovelAi官网中，使用{}增加权重x1.05。

5. Prompt支持使用emoji，且表现力较好，可通过添加emoji达到表现效果。如😍形容表情，🖐可修手。

6.“+” ， “ AND” ， “|” 用法：

“+”和“ AND ”都是用于连接短Tag，但AND两端要加空格。"+"约等于" and "

“|” 为循环绘制符号（融合符号）

(Prompt A: w1)|(Prompt B: w2)

以上表达适用于WebUI，w1、w2为权重。AI会对A、 B两Prompt进行循环绘制。可往后无限加入Prompt。

下面通过发色来测试：

(red hair:1.1)+(yellow hair:1.25)+(green hair:1.4)

(red hair:1.1)+(yellow hair:1.25)+(green hair:1.6)

(red hair:1.3)+(yellow hair:1.25)+(green hair:1.4)

固定SEED的简单彩发测试：

从左到右依次为

(red hair:1.2)+(yellow hair:1.4)+(green hair:1.4)

(red hair:1.2) AND (yellow hair:1.4) AND (green hair:1.4)

(red hair:1.2),(yellow hair:1.4),(green hair:1.4)

(red hair:1.2)_(yellow hair:1.4)_(green hair:1.4)

(red hair:1.2)|(yellow hair:1.4)|(green hair:1.4)

(red hair:1.2)(yellow hair:1.4)(green hair:1.4)

(red hair:1.2) (yellow hair:1.4) (green hair:1.4) Tag中间只夹空格效果也还不错

结论：

① (Prompt A） AND (Prompt B），DDIM采样器不支持且测试效果不太好，暂不推荐使用

② 目前来看，+号约等于“ and ”，其他用法也都可以尝试效果差不多。

7. 权重赋予

(Prompt A:1.5,Prompt B:1.5) 效果并不好，不如直接(Prompt A:1.5），（Prompt B:1.5)，彩发可以直接(red hair:1.2),(yellow hair:1.4),(green hair:1.4)

8. 分布渲染

句式：

[Prompt A:Prompt B:Step]

[Prompt A::Step]

[:Prompt B:Step]

此处Step > 1时表示该组合在前多少步时做为Prompt A渲染，之后作为Prompt B渲染。Step < 1时表示迭代步数百分比。

Prompt格式优化：

1. 简易换行三段式表达：

第一段： 画质tag，画风tag

第二段：画面主体，主体强调，主体细节概括。（主体可以是人、事、物、景）画面核心内容

第三段：画面场景细节，或人物细节，embedding tag。画面细节内容

第二段一般提供人数，人物主要特征，主要动作（一般置于人物之前），物体主要特征，主景或景色框架等

2.元素同典调整版语法：

质量词→前置画风→前置镜头效果→前置光照效果→（带描述的人或物AND人或物的次要描述AND镜头效果和光照）*系数→全局光照效果→全局镜头效果→画风滤镜（embedding）

Tips: 不同模型（ckpt）对Prompt的敏感程度不同，一套完善的prompt在不同的模型中，表达效果会有差异。每个模型都有自己的特色，需根据模型特色，慢慢调试prompt组合。

实例展示1:

parameters:

Prompt：

(((masterpiece))), best quality, ultra-detailed, extremely detailed CG unity 8k wallpaper,best illustration, an extremely delicate and beautiful,floating,high resolution,

dynamic angle,dynamic pose,(1girl),blue eyes,(multicolored_hair+silver hair:1.3+red hair:1.2+purple hair+yellow hair:1.3+green hair:1.3), white wuxia clothes,neck ribbon, beautiful face,

too many drops of water,cloud,twilight, wide shot,watercolor,

Negative prompt:

lowres, bad anatomy, bad hands, text,error, missing fIngers,extra digt ,fewer digits,cropped, worst quality ,low quality,normal quality, jpeg artifacts,signature,watermark, username, blurry, bad feet,fused body,

Steps: 34, Sampler: DPM++ 2S a, CFG scale: 10.5, Seed: 1351767725, Size: 960x576, Model hash: 925997e9, Eta: 0.68, Clip skip: 2, ENSD: 31337

GFPGAN visibility:1.Upscale: 3, visibility: 1.0, model:ESRGAN_4x

不建议在括号内多重赋予权重

实例展示2:

parameters:

Prompt：

(extremely detailed CG unity 8k wallpaper,masterpiece, best quality, ultra-detailed),(best illumination, best shadow, an extremely delicate and beautiful),floating,

[(((1girl)), hairband,earings,detailed and beautiful red eyes,naughty face,black neck ribbon,pink uniform,off shoulder,(doggy style),stockings,((blue:1.3)|(pink) long hair),cleavage,)::0.8]

forest,depth of field,blue splashing,glorious,petals,wind,leaves,grasslands, Volumetric light,

Negative prompt:

blush,lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name,bad feet,big head,fused body,multiple hands,multiple legs,multiple hands,multiple breast,multiple lower legs,

Steps: 149, Sampler: Euler, CFG scale: 11, Seed: 3374835118, Size: 1024x512, Model hash: 925997e9, Eta: 0.68, Clip skip: 2, ENSD: 31337

实例展示3:

parameters:

Prompt：

(extremely detailed CG unity 8k wallpaper,masterpiece, best quality, ultra-detailed),(best illumination, best shadow, an extremely delicate and beautiful),floating,high saturation,

[(dynamic angle,((1girl)),gorgeous,(long blue hair)|(long pink hair),hairband,beautiful earings,delicate and beautiful red eyes,glasses,open mouth, neck ribbon,pink_sweater ,off shoulder,bra (strap),dynamic pose,delicate white stockings,(cleavage),panties,looking at viewer,long shot): :0.75],

[ :(colorful forest,depth of field,(many flying drops of water:1.2),petals,toys around,wind,leaves,grasslands,Volumetric light):0.4]

Negative prompt:

blush,lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts,signature, watermark, username, blurry, artist name,bad feet,big head,fused body,multiple hands,multiple legs,multiple hands,multiple breast,multiple lower legs,multiple thighs,fused body,twist body,

Steps: 150, Sampler: Euler, CFG scale: 11, Seed: 2949839298, Size: 1024x576, Model hash: 925997e9, Eta: 0.68, Clip skip: 2, ENSD: 31337

多人物实例：

自实验格式：

(extremely detailed CG unity 8k wallpaper,masterpiece, best quality, ultra-detailed),(best illumination, best shadow, an extremely delicate and beautiful),floating,high saturation,

((2girls)),dynamic angle,[girlA+(特征),girlB+(特征): :0.5],

[ :(colorful forest,depth of field,(many flying drops of water:1.2),petals,toys around,wind,leaves,grasslands,Volumetric light):0.5]

也有一定概率能出

一些有趣的实例：

Prompt：

Kawaii low poly Animal（替换词） character, 3d isometric render, white background, ambient occlusion, unity engine, square image

Negative Prompt：

nsfw, lowres, bad anatomy, bad hands, text, error, missing fingers,extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry,

Steps: 70, Sampler: Euler, CFG scale: 10, Seed: 720579354, Size: 512x512, Model hash: 81761151, Eta: 0.68, Clip skip: 2, ENSD: 31337