关于 AUTOMATIC1111/stable-diffusion-webui 的 FAQ

该 FAQ 并非意图一次性提供所有必要信息, 仅仅是提供必要的指路. 最后的杂项部分是宝藏区, 建议都翻翻看啦!

查看最新消息参阅 SD RESOURCE GOLDMINE (English), 或者 sudoskys/StableDiffusionBook (中文)

前言

建议优先阅读官方文档或者 SD RESOURCE GOLDMINE 或者 VOLDY RETARD GUIDE或者としあきdiffusion

中文教程见 sudoskys/StableDiffusionBook.

如果以上教程都不想看, 为什么不试试这个超小白教程, 或者听听暴躁老哥的教程 呢?

我假设读者已经搭建好了 AUTOMATIC1111 分支的 stable-diffusion-webui.
若尚未搭建完成请参照其他教程, 亦可参照 Troubleshooting 一节

注: 使用 NAIFU (NovelAI AI Image Frontend Ultimate) 或者 NovelAI 官方服务的读者会可能会发觉此文没什么卵用, 本人也不推荐使用基于
泄露版 NovelAI 的前端, 因其不再更新, 且缺少必要的 Prompt Mixing 语法, 若有需求可参照 NovelAI 的 Prompt Mixing 一节.

效果不如 NovelAI 官方服务

请仔细阅读该 disccusion Emulate NovelAI

不觉得官方的画风太无趣了么?)

negative prompt, 加载 hypernetworkvae. (见 Hypernetwork 一节)

Stop At last layers of CLIP model ~ignore last layers of clip mode~ 改成 2: 不能套其他模型, 这是 NovelAI 专属 hack (commit)

甚至我会说Novel AI模型skip ~0~ 1 都是玄学,只有2才是正确食用方式

prompt 加上 masterpiece, best quality 有奇效, 见

还记得泄露的文件夹里边有个 config.yaml 么? 将其改名为 模型前缀.yaml 和模型丢在一起就能加载啦, 效果还是有提升的. (commit)

v2.pt 我个人认为不够 interesting, 不做展开介绍, 参阅 Emulate NovelAI (I don't like it personnally).

NovelAI 默认 Negative Prompt source

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

Emulate NovelAI 里面还给出了其他 hacks, 像是改 eta (What are "Sampler Parameters"?)

NAI preset seed: 31337 emulates strenght:0.69, noise: 0.667.

其他模型

别抱着 NovelAI 泄露的模型不放, 那玩意虽然 consistent 但是结果看多了挺无趣的, ~为什么不试试混合 (merge) 模型呢? (have a try but don't be too serious)~

Fine Tune

  • 原模型认识的概念就 Textual Inversion (人物风格均可)
  • 风格化就炼 Hypernetwork (我会说 Hypernetwork 仅仅是把原模型认识的画风更加具象地表达出来了, 没有增添任何的新东西)
  • 不认识的概念就 DreamBooth (我愿称之为 Copy and Paste Machine, 使用时请保证数据集 consistent)

<details>
<summary>更多</summary>
<p>请参照其他教程, 本 FAQ 并非专注于 Training. </p>
</details>

Embeddings/Textual Inversion

Textual Inversion

丢到 embeddings 文件夹, prompt 提到你要用的 embedding 的文件名即可

reports that training with medvram is possible.

想要找现成的 TI embedding? 看看 list of Textual Inversion embeddings for SD. 请自行寻找.

Unload VAE and CLIP from VRAM when training 这个选项完全无用, 仅仅是把 VAE 加载到内存而已节省显存罢了, 要真正去掉 VAE 目前只能改名. 训练 Embedding 的时
必须卸载 VAE, 否则会出现错误的生成效果.

给出两组参考 Learning Rate, 请自行调整

  • 1e-4:2000, 3e-5:5000, 1e-5:10000, 3e-6:20000, 1e-6
  • 5e-3:100, 1e-4:2500, 1e-5:20000, 1e-6

Hypernetwork

Move Hypernetworks and Model selection underneath "Generate"

你的 models 文件夹应该是这样的

./models
├── Codeformer
├── ESRGAN
├── GFPGAN
├── hypernetworks
│   ├── aini.pt
│   ├── anime_2.pt
│   ├── anime_3.pt
│   ├── anime.pt
│   ├── furry_2.pt
│   ├── furry_3.pt
│   ├── furry_kemono.pt
│   ├── furry_protogen.pt
│   ├── furry.pt
│   ├── furry_scalie.pt
│   ├── furry_transformation.pt
│   └── pony.pt
├── LDSR
├── Stable-diffusion
│   ├── final-pruned.ckpt -> novelai 的 model.ckpt (pruned)
│   ├── final-pruned.vae.pt -> novelai 的 animevae.pt
│   ├── final-pruned.yaml -> model.ckpt 同文件夹的 config.yaml
│   └── 其他模型文件.ckpt
└── SwinIR

启动 cli 有提示加载就 OK, 去设置选模型那里选喜欢的 hypernetwork.

可以指定不同阶段时的 learning rate 了 commit

训练 Hypernetwork 需要极低的 Learning Rate, 这里给出一组参考值 5e-5:100, 1e-5:1000,5e-6:2500, 1e-6:5000, 2.5e-6:10000, 5e-7.

Hypernetwork 不需要卸载 VAE, 即使卸载了也没什么问题. 卸载 VAE 正确的方式是修改文件名.

DreamBooth

TODO. Help wanted. Haven't used it myself.

参数解释

  • step: 迭代多少次, 取值和 sampling method 有关, DDIM 收敛较快, 见 Sampler vs. Steps Comparison (low to mid step counts) 我用 DDIM 取 ~20 够了~ 18 足矣.
  • batch count/batch size: 决定生成的图片数量, 显存够就加 batch size, 不够就 batch count, 得到的图片数量是两者之积 (小显存还是只动 count 就好)
  • sample method: 我无脑 DDIM, Eula 也挺好用? (带 a 的是 ancestral 的意思, step 增长出图不稳定)
  • cfg scale: 符合 prompt 的程度, 值越高越会字面看待 prompt, 低则给模型较大的发挥空间, 但是实际模型表现上来看 cfg scale 低 (6-8) 饱和度低, 偏线稿, 偏杂乱, 高 (18-22) 则饱和度偏高, 偏 CG 风格.
  • denoise strength: img2img 专属参数, 从 0 到 1 取值, 值越高 AI 对原图的参考程度就越低 (同时增加迭代次数), 我个人喜欢低 cfg 高 denoise 重绘图, 高 cfg 低 denoise 改细节.

Euler a (short for ancestral) produces great variety with a low number of steps, but it's very difficult to make small adjustments. The non-ancestral samplers all produce mostly the same images as the number of steps increases, use Euler if you're unsure. (source)

Can anyone offer a little guidance on the different Samplers?

Sampling method selection

关于 cfg 和 denoise strength 有张好图可以说明. 原图为Lenna. (XY plot 大图分流)

denoise strength and cfg

不过这些参数都能自己用 XY plot 跑

Extension

See Extension on wiki.

Aesthetic Gradients

原本为已经 merged 的 PR, 被下放到 Extension 中.

Aesthetic Gradients

See Implementation of Stable Diffusion with Aesthetic Gradients (Pull Request).

Also chec Custom Styles in Stable Diffusion, Without Retraining or High Computing Resources

I have no idea how to train this and it is NOT interesting enough.

Wildcard

原本为 Custom Script, 被下放到 Extension 中.

/scripts/wildcards/ 文件夹, 建立命名为 idkwhatever.txt (显然 idkwhatever 可以被替换成任意词语), 放你想要替换的 token, 一行一个.

在 Custom Script 中起用 wildcards, prompt 中填写 __idkwhatever__ 便会每次随机 roll idkwhatever.txt 中的内容.

搭配 Generate Forever 效果更佳.

不知道怎么写? 看看Wildcard Lists.

Image browser

集成图片浏览器, 原本为内置功能, 被下放到了 Extension 中.

Custom Script

在 Extension 出现之前就有的自定义脚本.

Custom Scripts

放到 scripts 文件夹下, 加上 --allow-code 参数启动.

Performance/Others

xformers

xformers, 加速推理, 对质量影响未知, 分辨率越高加速效果越好, 见 What does the new xformers update mean?

在 Windows 上编译似乎比较麻烦 (I'm running Linux anyway) Windows version installs binaries mainained by C43H66N12O12S2. (only for Ampare) (source)

加上 --xformers 参数即可, 应该会自动装上构建好的 wheel, 但是也可以自行编译.

Aside from your wheel not working with older architectures, is there a reason to drop pre-Ampere support? No. (source)

我自己是用 pip install -U setuptools 装最新版 setuptools 才编译成功的 (python 3.10)

Image 逆向 Prompt

  • Interrogate CLIP (在 img2img 标签中, 出来的描述偏 SD 原版自然语言)
  • DeepDanbooru 审问出来的是 tags. 加上 --deepdanbooru 作为启动参数 (commit)
  • LenKiMo_Bot 闭源, 后端未知, 不过效果不错

高分辨率下出怪图

我原本以为是低显存触发 split attention 会这样, 不过似乎不管显存高低高分辨率都容易出怪东西.

群友的修复方法 (source)

  1. 使用 —medvram 或者 —lowvram 参数启动webui
  2. 正常输入 prompt,选择较小分辨率生成图片。记住你生成图片的分辨率。生成完毕之后,复制图片的Seed。
  3. 重点来了!生成完毕后,先查看图片是否满意。如果效果满意,直接将图片扔进Img2img。可点击“Send to img2img”。
  4. 在img2img界面底部,有一 Script 选项。将 Script 选为 SD Upscale,里面的 Tile overlap 尽量调小。
  5. 一般地,扔进Img2img的图,会继承原txt2img的所有prompt和negative prompt。如果你发现prompt有变动,回到txt2img,复制prompt到img2img。
  6. 选择合适的Sampling Steps和Sampling method。
  7. 确认你的 Width 和 Height 与原图一致。这一步非常重要!这里的 Width 和 Height 是超分时 img2img 的图片分割大小,如果不合适会导致出现重叠问题。
  8. 将第 2 步复制的 Seed 填入img2img的 Seed 里。
  9. Generate!

或者用 highres-fix

By default, txt2img makes horrible images at very high resolutions, and this makes it possible to avoid using the small picture's composition.

使用CPU进行推理

--use-cpu all 尽可能的使用CPU进行生成 (PR)

多 GPU

临时解决办法, 开多个实例, 每个实例指定 GPU.

Notes/Tips

如何写 Prompt

AI 画图要是还要 AI 生成 Prompt 那身为人类的我们不就是只会点击鼠标的工具么?

TODO. See SD GUIDE FOR ARTISTS AND NON-ARTISTS IN-DEPTH TIPS, TRICKS, TUTORIALS AND MORE. What I would do is just copy and paste/translate this document.

下面将介绍一些高级语法来帮助您更好地控制生成的图像.

Tips

  • 逗号用得越少越好, 除非要明确不同主体

增强 prompt

Attention/emphasis

NovelAI 语法不同, 用 () 增强, 减弱仍然是 [] (但强度不同, AUTOMATIC1111 为 1.1倍, NovelAI 为 1.05 倍)

1
2
3
a (word:1.5)  # token 强度乘上 1.5  (增强) 约为 ((((word)))) (1.1^4 = 1.4611)
a (word:0.25) # token 强度乘上 0.25 (减弱) 约为 [[[[[[[[[[[[[[word]]]]]]]]]]]]]]]] (1/1.1)^14 = 0.263331254
a \(word\)    #  \ escape 掉括号, 用于你真的想用括号作为 token 的一部分的时候 (常见于人物)

Prompt Editing/Mixing

Prompt Editing 生成时实时调整 prompt 比重. 同时也能起到低 token 数的同时使得画面更有多样性.

1
2
3
4
# when 小于 1 是百分比, 大于 1 是确切的 step 数目
[from:to:when] # 从 from 来, 到 to 去, 于 when 时开始
[to:when]      # 从 when 时加上 to   (一开始没有该 token)
[from::when]   # 从 when 时拿去 from (一开始有该 token)

示例 prompt 透明雨衣

1
2
3
[lewd:assertive:0.3] small sweat [nude::0.8] [nipple::0.95] 1girl with disheveled messy atmosphere cherry bloom hair sitting in white [((((translucent)))):((transparent)):0.9] [hoodie:trench coat :0.4] liquid crystal [((transparent)) sailor dress::0.7] indoor inside [vehicle], [downpour raining:0.5] [nightclub:cyberpunk:0.3] city, public crowd

Negative prompt: ugly, fat, obese, chubby, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands fingers), messy drawing, morbid, mutilated, tranny, trans, trannsexual, [out of frame], (bad proportions), (poorly drawn body), (poorly drawn legs), worst quality, low quality, normal quality, text, censored, gown, latex, pencil

~PS: 我在想有没有类似 {word:step_start:step_end} 的等价替代, 可以让更灵活地替换 prompt.~

import modules.prompt_parser
modules.prompt_parser.get_learned_conditioning_prompt_schedules(['abc [:[x::0.3]:0.17] def'],100)
Out[53]:
[[
  [17, 'abc  def'],
  [30, 'abc x def'],
  [100, 'abc  def']
]]

#  0-16 'abc  def'
#  17-29 'abc x def'
#  30-100 'abc  def'

[:[word::end]:start] 是中间某步加 prompt 的正确语法 ~(真他娘怪)~

Alternative

alternate prompt

奇数步与偶数步交替 source (或者循环每一项, 看你怎么理解)

# photo of [cow|horse|bird]
# translated as

photo of cow
photo of horse
photo of bird
photo of cow
photo of horse
photo of bird
...              # you know, Alternative
photo of cow
photo of horse
photo of bird

怪东西, ~在想有什么更加实际的使用例子~. [messy sketch|watercolor] 或者 [sketch|oil paint] 可以混合画风? heterochromia [left blue|right red] eyes 提高异色瞳概率...?

NovelAI 的 Prompt Mixing

NovelAI 的思路截然不同, 它并不是指定加入或者去除 token 的步数, 而是对 整个 prompt 进行融合, 注意是整个完整的 prompt (理论上也可以用在单个或者部分 token 但是他们没这么写), 可能是懒得写分词 (加个括号的事情).

权重可以取 0.1100 (why?). 权重可为负数但 -1 时输出黑色图.

prompt_a:weight_a|prompt_b:weight_b|prompt_c:weight_c 
# 以此类推
# 完整 prompt 见个人 prompt 的枪娘部分
pointing at viewer:0.1 | upper body {{{fang}}} 1girl  aiming at viewer holding handgun {{muzzle flash}} in black serafuku red ribbon [[black leather gloves]] {dual wielding}, cyberpunk city

NovelAI 并没有 AUTOMATIC1111 版本的 Prompt Editing 的完全等价替代, 不过这也已经能增加一部分多样性了. (仍然不够灵活, 我希望可以将两者结合起来)

具体实现见源码 (日后官方可能会有改动)

Composable Diffusion

所谓 AND 语法, 暂时没研究能怎么用, undocumented.

Composable-Diffusion Compositional Visual Generation with Composable Diffusion Models

See commit.

Wait for Implement multi-cond guidance for Composable Diffusion to merge. Not gonna happen.

Negative Prompt 十分重要

部分 negative prompt 有做训练时数据集标注, 像是 bad anatomybad hands 以及 error 这些在 Danbooru 都有实际标记的 tag, 故请使用那些正确的 tag 以获得更好的效果.

其他提高质量的 negative token 属于玄学范畴. 我会说 negative propmt 的提升和 positive prompt 一样甚至更加重要. 同样的一个正面 prompt 在有和没有 negative prompt 生成的效果有天壤之别, 所以请不要吝啬. 但是也不要走向另一个极端叠 buff, 这会让你的图像看起来 less interesting.

有意思的玩法

个人用 Prompt

注意这些 prompt 只能在 AUTOMATIC1111 的 WebUI 中使用 (因为用了 mixing 语法) 请勿直接丢到 NovelAI 出图, 质量不做保证.

肏桌子

solo seductive small sweat blush 1girl masturbation against [desk] standing (((orgasm))) opalescent eyes with disheveled messy white [rainbow] streaked hair [in open shirt:with white pantie:0.5] in indoors, hydrophobic wet cream dripping [pussy] ejaculation squirting, presenting, looking at viewer, fluorescent cyberpunk city, radiant [infrared:ultraviolet:0.5] blacklight accent lighting
Negative prompt: ugly, fat, obese, chubby, (((deformed))), [blurry], bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing, morbid, mutilated, tranny, trans, trannsexual, [out of frame], (bad proportions), (poorly drawn body), (poorly drawn legs), worst quality, low quality, normal quality, text, censored, sitting, chair, big large medium huge breasts, full

枪娘

1
2
3
4
5
upper body (fang:1.1) 1girl [(pointing at viewer:1.3):(aiming at viewer:1.5):0.38] (muzzle flash:1.1) (holding:1) [(handgun:1.1):(revolver:1.6):0.8] in black serafuku red ribbon (black leather gloves:0.89) dual wielding, floating bullet, cyberpunk city
Negative prompt: bad anatomy, bad hands, text, missing fingers, signature, watermark, blurry, deformed, [blurry], bad anatomy, disfigured, poorly drawn face, mask, glasses, (mouth mask), camera

# NovelAI 语法 (见上文 prompt mixing) (感觉这个出图更 consistent)
pointing at viewer:0.1 | upper body {{{fang}}} 1girl aiming at viewer holding handgun {{muzzle flash}} in black serafuku red ribbon [[black leather gloves]] {dual wielding}, floating bullet, cyberpunk city

透明雨衣, 见 Prompt Editing/Mixing 一节

VR 眼镜

sweat flat nude (fang) 1girl in (vr headset) sitting [spread legs:m legs:0.6] masturbation (orgasm) [shake hips] ((with machine)) white thighhighs, [(dildo): (viberator):0.6], party ((in front of a large crowd)), (wet pussy), [white wet cream dripping fluid]
Negative prompt: lowres, bad anatomy, bad hands, text, missing fingers, worst quality, low quality, normal quality, signature, watermark, blurry, eyes

吸烟

1
2
3
# ((飞机耳)) 加个玄学 token, 虽然没有兽耳
[(:o:0.5):0.3] 1girl [(hand on hips:0.8):(hand in pockets:1):0.35] standing smoking in black serafuku pleated skirt red ribbon, fire lighter, [pocky :0.05] cigarette in mouth, steam, 3/4
Negative prompt: lowres, bad anatomy, bad hands, text, error, cropped, worst quality, low quality, normal quality, blurry, deformed, ugly, full, long fingers, multiple

章鱼娘

1
2
3
4
# 注: イカ 是章鱼, 而 イカ耳 却是飞机耳 (why?) 
# 这展现了 SD 是可以理解多语言的 (理解了, 又没完全理解)
(イカ耳:1.8) cat ears 1girl toast in mouth late for school
Negative prompt: lowres, bad anatomy, bad hands, text, error, cropped, worst quality, low quality, normal quality, blurry, deformed, ugly, full, long fingers

精液浴

1
2
3
4
5
6
# 前两天一直有人在问就来调一下, 怎么那么简单的东西怎么都调不出来. 
# seman Stable Diffusion 不认, cum 认但是只会到身体或者少量滴在地板, 唯一的思路就是 fluid, 很多很多的 fluid
nude 1girl [reclining:bath:0.6] [urinal:bathtub:0.6] in (wet cream white [translucent:0.6] [goo:fluid:0.4]:1.5), cum on body face , (wet cream dripping pussy:1.2)
Negative prompt: lowres, bad anatomy, bad hands, bad legs, text, error, cropped, worst quality, extra arms, ((extra legs)), low quality, normal quality, blurry, deformed, ugly, full, latex, bodysuit, leather

(Sampler: DDIM, CFG scale: 6), Model hash: e6e8e1fc, Clip skip: 2

白浊液喷尿高潮 (比上面的更色了)

1
2
3
4
5
# 不要害怕 token 重复, 你觉得某样东西很重要肯定要重复, 比如说那黄色的圣水
(nude:1.1) sweat sweatdrop small breasts nipples 1girl [reclining:bath:0.8] [(urinal:1.1):(bathtub:1.1):0.65] (orgasm:1.2) (peeing:1.2) (female ejaculation:1.1) (pee:1.4) in [(wet cream white [translucent:0.6] [goo:fluid:0.5]:1.5):((golden:1.3) [translucent:transparent:0.75] pee fluid oil:1.5):0.7], [(golden:1.3) water wet [translucent:transparent:0.8] pee fluid fluidity oil:1.1):0.7], cum on body face, female ejaculation (wet cream dripping pussy:1.2)
Negative prompt: lowres, bad anatomy, bad hands, bad legs, text, error, cropped, worst quality, extra arms, ((extra legs)), low quality, normal quality, blurry, deformed, ugly, full, latex, bodysuit, leather, silk, futa, trans, trannsexual, [[[dress]]], [[[pantie]]], [[clothes]], [[towel]], [[[[slime]]]]

(Sampler: DDIM, CFG scale: 6), Model hash: e6e8e1fc, Clip skip: 2

掀裙放尿

1
2
3
4
5
6
7
# 金色大厅版本 (写错语法了但是效果意外不错?)
disheveled messy (fang:0.9) [long hair:0.9] [(two side up:1.1)::0.8] flat breasts 1girl [with heterochromia multicolored red eyes::0.8] in black red ribbon sailor dress (white socks:0.9) (peeing:1.4) standing on liquid female [(ejaculation:0.6)::0.9] orgasm, ([lifted by self:lift dress:0.4]:1.3), (wet cream dripping [cum in::0.9] pussy:1), [(dripping golden (transparent:0.7) pee (oil:0.4) water fluid :1.6) from pussy::0.85], (golden fluid [on the ground::0.9]:0.6), no panties, [American shot::0.3], [cowboy shot::0.35]
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, blurry, comic, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], [2girls], [yuri], [sitting], lying, ((closeup)), lower body, large breasts, (breasts out:1), gown, latex, pencil

# 正常放尿
disheveled messy (fang:0.9) [long hair::0.9] [(two side up:1.2):0.7] flat breasts 1girl [with heterochromia multicolored red eyes:0.8] in black red ribbon sailor dress (white socks:0.9) (peeing:1.4) standing on liquid female orgasm, ([lifted by self:lift dress:0.4]:1.3), (wet cream dripping [cum in:0.8] pussy:1), [(dripping golden (transparent:0.7) pee (oil:0.4) water fluid :1.6) from pussy:0.8], (golden fluid [on the ground::0.9]:0.6), no panties, [American shot::0.3], [cowboy shot::0.35]
Negative prompt: bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, blurry, comic, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], [2girls], [yuri], [sitting], lying, ((closeup)), lower body, large breasts, (breasts out:1), gown, latex, pencil

咬裙子, 超级涩!

# 感谢群内大佬 @TLMEMO 的 prompt, 我愿称您为顶级魔法师
# 原始 prompt
# masterpiece,best quality,solo a girl lift her dress on mouth and use mouth bite the dress,clothes in mouth,navel,pussy

# 金色大厅版本 (写错语法了)
dishevled messy [long hair:0.9] [(two side up:1.1)::0.6] small breasts 1girl [with heterochromia multicolored red eyes::0.8] lift her [long shirt:dress:0.4] on mouth and use mouth bite the [long shirt:dress:0.4] standing on liquid female [(ejaculation:0.6)::0.9] , [mouth hold:(dress lift:0.1) with mouth:0.7], (wet cream dripping [cum in::0.9] pussy:1), [(dripping golden (transparent:0.7) pee water fluid :1) from pussy::0.85], (golden fluid:0.6), breasts out nipples, navel, [(panty pull:0.6)::0.8], [American shot::0.3], [cowboy shot::0.35]
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, blurry, comic, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], [2girls], [yuri], ((closeup)), lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit, vomiting, (handkerchief:0.8), sandwich, (cum in mouth:1.2)

# 修正后
dishevled (:3:0.1)messy [long hair:0.9] [(two side up:1.1)::0.6] small breasts 1girl [with heterochromia multicolored red eyes:0.8] lift her [long shirt:dress:0.4] on mouth and use mouth bite the [long shirt:dress:0.4] standing on liquid female [(ejaculation:1):0.8] , [mouth hold:(dress lift:0.1) with mouth:0.7], (wet cream dripping [cum in::0.9] pussy:1), [(dripping golden (transparent:0.7) pee water fluid :1.6) [between legs:from pussy:0.95]:0.9], (golden fluid:0.6), breasts out nipples, navel, no panties, [American shot::0.3], [cowboy shot::0.35]

# 无放尿版 (更加 consistent?)
dishevled messy [long hair:0.9] [(two side up:1.1)::0.6] small breasts 1girl [with heterochromia multicolored red eyes::0.8] lift her [long shirt:dress:0.3] on mouth and use mouth bite the [long shirt:dress:0.3] standing on liquid female [(ejaculation:0.8)::0.8] , [teeth hold:(dress lift:0.1) with mouth:0.7], (wet cream dripping (cumdrip:1.2) [cum in::0.9] pussy:1), (golden fluid:0.6), breasts out nipples, navel, [(panty pull:0.6)::0.8], no panties,[American shot::0.3], [cowboy shot::0.35]
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, blurry, comic, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], [2girls], [yuri], ((closeup)), lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit, vomiting, (handkerchief:0.8), sandwich, no pantie, (cum in mouth:1.2), (gag:0.8), fellatio

工口蹲涩猫猫

1
2
3
4
5
# 用于 benchmark hypernet/embedding 或者是某 artist 的效果
# 最后的 [messy sketch|watercolor] 可酌情删去
cat ears [lewd:assertive:0.2] small breasts [(loli:0.9):0.8] (fang:1) nude girl with in in bright silver disheveled (two side up:1) messy disheveled (long:0.9) hair with hair flower ornament squatting (double middle finger:1.1) [hands up:0.3], [female orgasm:(female ejaculation:1):0.5], (wet [cum in:0.8] cumdrip pussy:1), cowboy shot, [messy sketch|watercolor]
Negative prompt: lowres, (bad anatomy), bad leg, bad proportions, bad perspective, signature, watermark, username, text, worst quality, low quality, blurry, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], lower body, large breasts, censored, (blush:1), panties
Steps: 18, Sampler: DDIM, CFG scale: 11, Size: 384x640, Model hash: 925997e9, Batch size: 6, Batch pos: 0, Clip skip: 2

两百一次雌ガキ

[(sketch:0.95)|illustration|(oil painting:1.2)] of [machinery::0.1] [(evil grin:0.85)::0.5] [fang:0.5] flat breasts pervert [lewd|assertive] sweat [(loli:0.7)::0.6] 1girl (presenting:1.1) long two side up hair nude squatting looking at viewer [hands up::0.6] [(v:1.22) gesture|(victory signs:1.2):(peace sign:1)] [spread legs:m legs:0.6] with ([sex:toy:vibrator:0.6]:0.95) [with machine::0.2], cumdrip uncensored (spread:0.6) pussy (clitoris:1.25) (labia minora:1.25), navel, nipple with (vibrator:1.2), female orgasm ejaculation, cowboy shot
Negative prompt: lowres, bad anatomy, (bad hands:0.25), missing fingers, (extra digit:0.5), (fewer digits:0.5), bad leg, bad proportions, bad perspective, text, worst quality, low quality, normal quality, blurry, emphasis lines, futanari, futa, erection, 1boy, multiple girls, missing limb, extra legs, [mutation], lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit, (high heel), censored, monochrome, (blush:0.7), [(sex:0.5):sex:0.4], [comic::0.4], 2koma, 3koma, 4koma, instant loss, collage, multiple girls, [pov::0.25]

猫猫露出

1
2
3
4
5
6
7
# 分辨率宽屏注意
# 建议搭配 aki99 或者 Nachoneko 的 hypernetwork, kinnnotama 也行
# 适当调节权重, 更换姿势 (像是 [spread legs|m legs] 之类的常规操作)
# NovelAI 还是很懂 21:9 的构图的
(closeup:0.5) [(loli:1):(loli:1.1):0.8] [nude:0.7] flat chest (small breasts:0.9) nipples cat ears 1gir with two side up long curly messy disheveled hair standing in shirt on [realistic::0.2] in cyberpunk neon (night:1.05) [raining:0.8] city public (crowd:0.95) (masturbation:1.1) exhibitionism presenting, flower hair ornament, cowboy shot, (wet:1) (cumdrip pussy:1), (collarbone:1.05), [sketch|line art] [(oil painting:1.2):0.7] spot color by [(Jeremy Mann:1.4)::0.6] [(Akihiko Yoshida:1.4):0.4], [dynamic angle:0.7]
Negative prompt: bad legs, bad face, bad eyes, 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, elf, panties, pants, 1koma, 2koma, 3koma, 4koma, comic, male, boy, fellatio, [sex:0.7], doggystyle , vomit
Steps: 21, Sampler: DDIM, CFG scale: 11, Size: 1280x576, Model hash: 925997e9, Hypernet: Nachoneko-hn, Batch size: 4, Batch pos: 2, Denoising strength: 0.7, Clip skip: 2, First pass size: 896x384

机械猫猫露出

# 只要把手砍掉 AI 就不会画不好手了 (认真)
# 说实在的, 换成义肢的效果是真的好, 增加了 SF 风又解决了手脚问题
# 以下的权重是推荐带 hypernet/embeddings 的
# 因为通常 hypernet/embedding 会强化人物表现, 所以权重的数值需要提升压下人物
# 建议配合的 himitsu hyper 或者我炼的 kantoku 线稿 embedding (见 https://t.me/StableDiffusion_CN/193551)
# TODO: 可靠的托管文件平台
# 21:9 分辨率注意
(closeup:0.5) [(loli:1):(loli:1.1):0.8] ([pervert|assertive]:0.9) dieselpunk machinery [nude:0.6] flat chest (small breasts:0.9) nipples cat ears 1girl with white  two side up long curly messy  disheveled hair reclining in (trench coat:1.05) on [realistic::0.2] in (night:1.05) [cyberpunk|dieselpunk] sci-fi public city (crowd:1) (masturbation:1) in white thighhighs exhibitionism presenting with ((mecha:0.8) cyborg (prosthesis:1.1) artificial organ:1.05), flower hair ornament, cowboy shot, navel (wet:1) (cumdrip spread pussy:1.1), (collarbone:1.05), [sketch|line art] [(oil painting:1.2):0.7] spot color by [(Jeremy Mann:1.45)::0.6], [dynamic angle::0.4]
Negative prompt: bad legs, bad face, bad eyes, 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, elf, panties, pants, 1koma, 2koma, 3koma, 4koma, comic, male, boy, fellatio, [sex:0.7], doggystyle , vomit, [gundam::0.7], [bodysuit, swimsuit::0.6], pencil, gown
Steps: 20, Sampler: DDIM, CFG scale: 11, Seed: 416762085, Size: 1280x576, Model hash: 925997e9, Batch size: 4, Batch pos: 3, Denoising strength: 0.7, Clip skip: 2, First pass size: 832x384

艺术家风格

原版的 Stable Diffusion 中便已经有许多 artists 的风格印在 "脑" 海里了 (见杂项中的 Artist Style Study 部分), 我们需要做的只是去认识并调用它而已.

当然我不是反对训练新模型, 新模型当然是有必要的, 不过当原始的 artists 风格要玩明白也够玩不短时间了.

Jeremy Mann sketch 风

核心词条为 by [ (Jeremy Mann:1.7) :(Jeremy Mann:1.4):0.6] [and (Patrice Murciano:1.3)::0.4] [and (Carne Griffiths:1.4):0.6], [messy sketch|watercolor] 这是个缝合怪, 所以结果也不稳定, 偶尔色彩绚丽 (Patrice Murciano 和 Carne Griffiths 所带来的), 但我主体仍然想强调一个 high-contrast 的风格, 类似摄影的黑白底片.

权重看着调, 不同动作和背景的权重效果不一样.

1
2
3
4
5
6
7
8
9
# 配合 anime 风格的 hypernetwork 有奇效 (需要拉高 artist 风格比重)
# 推荐甘城猫猫和毛玉老师的 hypernetwrok
cowboy shot (anime:1) (fang:0.8) cat ears assertive nude (loli:1) girl in bright silver disheveled (two side up:1) (long:0.9) hair (looking at viewer:0.9) with hair flower ornament sitting (presenting:1.1) female ejaculation in open wet [(wedding:1.1):0.7] [(queen:0.8):0.8] (sailor dress:1.2) red ribbon white thighhighs in (street:1.05) [crowd:0.8], (public indecency:1.05), sexually suggestive, (labia wet cream cumdrip dripping pussy and (anus:0.8):0.8) by [ (Jeremy Mann:1.7) :(Jeremy Mann:1.4):0.6] [and (Patrice Murciano:1.3)::0.4] [and (Carne Griffiths:1.4):0.6], [messy sketch|watercolor]
Negative prompt: lowres, (bad anatomy), (bad hands), bad leg, bad proportions, bad perspective, signature, watermark, username, text, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, blurry, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit, (high heel), realistic, (blush:0.6), (tongue out:0.6)

# 最原始的版本 
[cowboy shot:0.5] (anime:1.2) small breasts nipples assertive nude girl in (twintails:1) (long:0.9) hair looking at viewer standing (presenting:1.1) female ejaculation white thighhighs in street [crowd:0.8], public indecency, sexually suggestive, labia wet cream cumdrip dripping pussy and (anus:0.8) by [(Jeremy Mann:1.6):(Jeremy Mann:1):0.6] [and Patrice Murciano::0.4] [and Carne Griffiths:0.6], [sketch|watercolor]
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, blurry, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit, (high heel)
Steps: 18, Sampler: DDIM, CFG scale: 11, Seed: 1469635569, Size: 384x640, Model hash: 925997e9, Clip skip: 2

新川洋司 (Yoji Shinkawa)

1
2
3
4
5
# See also 
# by [(Yoshitaka Amano:1.8):(Yoshitaka Amano:1):0.7], ink sketch, spot color
[cowboy shot:0.5] small breasts (fang:0.5) assertive nude girl in multicolored rainbow (twintails:1) (long:0.9) hair looking at viewer (squatting:1) (presenting:1.1) female ejaculation white thighhighs, sexually suggestive, labia wet cream cumdrip dripping pussy and (anus:0.8) by [(Yoji Shinkawa:1.8):(Yoji Shinkawa:1):0.7], spot color, ink sketch
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, blurry, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit
Steps: 18, Sampler: DDIM, CFG scale: 8, Seed: 3323305485, Size: 384x640, Model hash: 925997e9, Clip skip: 2

Agnes Cecile

很绚丽的彩色风

1
2
3
[cowboy shot:0.5] small breasts (fang:0.5) assertive nude girl in multicolored rainbow (twintails:1.2) long hair looking at viewer [sitting:squatting:0.4] (presenting:1.1) female ejaculation white thighhighs in street [crowd:0.8], sexually suggestive, public indecency, labia wet cream cumdrip dripping pussy and (anus:0.8) by [(Agnes Cecile:1.65)::0.7]
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, (extra digit), (fewer digits), worst quality, low quality, normal quality, blurry, comic, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit, (elf:1.1), high heels, (blush:0.6)
Steps: 18, Sampler: DDIM, CFG scale: 7.5, Seed: 1810426879, Size: 384x640, Model hash: 925997e9, Clip skip: 2

浮世绘/日本风

Kitagawa Utamaro 和 Hokusai, Katsushika 都有效. 套上二次元的时候就是增加许多日本风格的元素 (虽然并没有马上变成浮世绘) 试着修改下权重吧.

1
2
3
[cowboy shot:0.5] small breasts (fang:0.5) assertive girl in multicolored rainbow (twintails:1.2) long hair looking at viewer [sitting:squatting:0.4] (presenting:1.1) female ejaculation in serafuku pleated skirt white thighhighs in street [crowd:0.8], sexually suggestive, public indecency, labia wet cream cumdrip dripping pussy and (anus:0.8) by (Katsushika Hokusai:1.6)
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, (extra digit), (fewer digits), worst quality, low quality, normal quality, blurry, comic, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit, (elf:1.1), high heels, (blush:0.6)
Steps: 18, Sampler: DDIM, CFG scale: 7.5, Seed: 2854368763, Size: 384x640, Model hash: 925997e9, Clip skip: 2

H.R.Giger

1
2
3
4
5
# giger 不必多说, 玩 midjourney 都知道的 artist
# 我会说是触手必备 tag? 虽然不是我的 XP
[cowboy shot:0.5] small breasts (fang:0.5) assertive girl in (twintails:1.2) long hair looking at viewer [spread legs:m legs:0.2] (presenting:1.1) female ejaculation in serafuku pleated skirt white thighhighs in street [crowd:0.8], sexually suggestive, public indecency, labia clitoris wet cream cumdrip dripping pussy and (anus:0.8) by (giger:1.35), (looking at viewer:1.1)
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, (extra digit), (fewer digits), worst quality, low quality, normal quality, blurry, comic, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, [extra hands arms], large breasts, gown, latex, pencil, bodysuit, swimsuit, (elf:1.1), high heels, (blush:0.6)
Steps: 18, Sampler: DDIM, CFG scale: 7.5, Seed: 1270237875, Size: 384x640, Model hash: 925997e9, Clip skip: 2

Patrice Murciano and Carne Griffiths Cyberpunk style

1
2
3
[cowboy shot:0.5] small breasts nipples [( :o :1.2):(parted lips:1):0.9] assertive nude girl in (twintails:1) (long:0.9) hair looking at viewer (squatting:1) (presenting:1.1) female ejaculation white thighhighs, sexually suggestive, cyberpunk neon city street, labia wet cream cumdrip dripping pussy and (anus:0.8) by [(Patrice Murciano:1.6):(Carne Griffiths:1.2):0.6] [:[and (Ryan Hewett:1.2)::0.9]:0.6], [sketch|oil paint], cinematic lighting
Negative prompt: lowres, bad anatomy, (bad hands), bad leg, bad proportions, bad perspective, text, missing fingers, extra digit, fewer digits, worst quality, low quality, normal quality, blurry, emphasis lines, futanari, futa, erection, 1boy, missing limb, extra legs, slime, [extra hands arms], [mutation], lower body, large breasts, gown, latex, pencil, bodysuit, swimsuit, (high heel), [(realistic:1.2):0.1], thick lips, lipstick, 3d
Steps: 18, Sampler: DDIM, CFG scale: 8, Seed: 11307460, Size: 384x640, Model hash: 925997e9, Batch size: 6, Batch pos: 0, Clip skip: 2

Miscellaneous

都是好东西

总结类 (大而全)

特定主题

AI 图库/搜索引擎

Prompt Reference

对 Prompt Engineering 有帮助的参考

列表/咒语书

Wiki

Danbooru 常用 Tags Group

Artist Style Stuies

未分类

匿名版

特定平台

其他

Troubleshooting

已经被其他文档收录的注意事项, 仅供存档用

显卡生成黑色/绿色图

16 系显卡不支持半精度/混合精度

Green or Black screen

启动参数加 --precision full --no-half, 要省显存还能加 --medvram

使用 VAE 之后偶尔出现黑图

--no-half-vae can solve this? haven't tested yet (commit)

git 报错

HTTP_PROXYHTTPS_PROXY 环境变量. 或者用 proxychains. 或者路由器跑代理. 或者把 git clone 的仓库源换成 huggingface 的

pip

设镜像或者挂代理, 推荐清华源

懒人包改 Git 仓库进行更新

感谢 huige233 的贡献.

第一步

打开你的懒人包目录

目录图片

第二步

在路径栏输入cmd

路径图片

第三步

在打开的cmd窗口内输入git pull

pull图片

若提示'git'不是内部或外部命令

git

请下载一个 git

然后再执行第三步的步骤

第四步(如果在第三步看到了error请看第四步以及下面的步骤)

先将error内提示的文件保存一份
然后依次在cmd内输入

1
2
3
git fetch --all
git reset --hard origin/master
git pull

完成后把保存的文件放回原本的地方, 完成更新

Edit
Pub: 11 Oct 2022 02:49 UTC
Edit: 25 Oct 2022 12:27 UTC
Views: 1516