Skip to content

Practical Guide

This short practical guide will give you a quick overview of how to adjust the parameters appropriately to achieve the desired effect.

Currently, Ai does not handle details and multiple characters well, and the light and shadow structure will be messed up if handled one by one. If retouching, it is better to redraw. But it is practical to draw backgrounds and do previews.

The core logic is to narrow the range of pre-selected data, adjust the balance between negative and positive prompts, and use part of the grammar to control the intensity of the prompt word effect and the process of effect.

Inspiration Table

Character Expression Hair Eyes Clothes State Posture Camera
Scene Wide selection Light and shadow selection Background subject Character event place
Restriction Painting Type Evaluation Restriction Associative Elements 3D/2D
Events Acronyms SFW/NSFW Producers Brands


Recommended use of terms contained in Danbooru


citing an image from the Japanese Wiki, author unknown

parameters explanation
extreme closeup face closeup
close up` Headshot
medium close up ID photo
medium shot half body
cowboy shot no legs
medium full shot no legs
full shot full body


|category|description|examples|processing| |:----:|:----:|:----:|:----:|:----:| |rough|not enough details, characters look like rice cakes|![Lack of sample images]|needs to be properly enhanced step| |waves|dense waves or patterns|![missing example image]|vectors are not "uniform", maybe the weights are set too high| |model|The training process is very abstract|![missing image]|Vector is not "uniform", maybe the alchemy is fried|! |Conflict|Flat anime paper in real world|![missing example image]|may be a mix of cue words from different subgenres or themes|! |Rainbow|output of confusing rainbow-colored images|![Missing example image]|may be the generation resolution is set too low. |! |Size|outputs distorted images at large sizes|![Missing example image]|Turn on the Highres Fix option or use the Small Resolution + SuperScore method|

Noise Reduction

Real-ESRGAN]( is recommended for noise reduction for anime illustrations.



Extras for hypersegmented images

WebUi's extras page has its own extras function, which can be used to enhance the image quality using models like ESRGAN_4x.

If you want to extrascape Anime style illustrations, please use the realesrgan or realcugan tools directly for extrascape.

???? tip "Direct link to related models"

The files are downloaded to the `SDwebUI folder, models`.

[LDSR](, file size is 1.9GB

[BSGRAN 4x](, file size 63.9M

[ESRGAN_4x]( with a file size of 63.8MB

[ScuNET GAN/PSNR](" to D:\stable-diffusio\models\ScuNET\ScuNET.pth), file size [SwinIR 4x](, file size **Highres fix What Upscaler should be used?** We recommend using `ESRGAN_4x`, for detail-oriented use `SD Upscaler`.

WebUi Guide to Assisted Drawing

Want to re-draw or add elements manually with WebUi? Here's a little guide.

First of all, the effect of IMG2IMG2 and Inpaint is completely different. If you don't want the style to change, don't choose IMG2IMG2.

To avoid image stretching, the sizes should all be as close to the original size as possible, choose Crop and resize which means crop and resize

Masked Content

The Masked content related setting determines the content placed into the masked area before the repair, it determines the initial reference content (origin is the content of the mask before blurring, and latent nosie is referenced by noise (many random pigment dots).


latent noise is what determines the Ai reference, while Denoising strength can be interpreted as the degree of unreference to the reference.

Changing elements

If we need to change the elements, we can key and paste them.

If we need to add elements, we can use PS to graft a hand to the character and let Ai touch it up, or we can graft the lower body of another work for a bust without a lower body and let AI touch it up.


Make changes based on the original image

Using Inpaint, the main scene is removal/replacement.

Start by tracing thin lines around the edges of the figure, then hitting the color block (if there are shadows, take bright colors or draw full shadows). Change the intensity to choose a lower denoising of around 0.3 (the lower the closer to the input image).

Then use Img2Img Inpaint + relevant cue words to fix, not satisfied can be changed again until satisfied. Then the image is realesrgan superscored to remove the image texture.


Grafted image

Use PS software to add and remove elements and then re-produce them. This can solve the problem of drawing hands.

Ai also accepts other finished images for grafting (example of application: solving the problem of lying down without lower body)


sketching + IMG2IMG, this is the secret ~

Wide painting single figure generation is best to play the sketch, for color painting, to determine the main body of the picture.

Multi-figure to determine the number of characters, it is best to use the draft / colored 3d arrangement + figure generation.

More than three people will be difficult to control the effect, the number of people greater than 6 in the image model is not estimated...

Perform palm restoration

Feed the image into inpaint, use roughly the same cue words, put the cue about the hand in front, set the noise reduction depending on how much you want it to change (to 0.25 or less if you just want the hand to be more intact), then keep the steps the same as CFG and txt2img gen.

Or just mask the hand, fix it at full resolution, greatly reduce the padding (it uses the surrounding pixels to create the context, but is only re-creating the hand) and only hint at hand issues (detailed hand depiction, etc.)

The higher the CFG, the more it matches the cue words, and the higher the noise reduction the more it deviates from the original image.

Identical Characters & Differences

We can use LoRA to train a style model of similar characters, then use VRoid to roughly reproduce the characters and export them to VRM format, then use UE5's Control Rig function to adjust their movements, and finally use the Img2Img method to generate these characters with the same appearance. 10

For expressions or backgrounds, the Redraw (part of Img2Img) technique from the advanced tutorial can be used.

  • ControlNet

ControlNet is a neural network structure to control diffusion models by adding extra conditions.

Iterative sketching 5

Here is a discussion of how to optimize hand sketches by Ai drawing, note not quadratic.

In the first iteration, not too many Steps are needed, CFG can be very low (for better diversification results), and Denoising should be around 0.3-0.4 if you don't want to lose the sketch completely.

In the final iteration, increase Steps and Denoising intensity (but not more than 0.8, otherwise the image will be corrupted, especially at greater than 512*512) see here, while increasing the CFG and size as needed.

You can always fix the hints (add or remove details that appear) and try a different sampler.

Also, shouldn't you use a fixed constant seed for initial generation?

If you provide a seed (instead of a random -1), your image will soon become oversaturated, oversharpened, and overpixelated. .... Of course you can use a fixed seed if you want to fine tune it.

Colour rather than lines

Whether it's 3D (3D models like DAZ) or line art, AI only recognizes color, not lines, and color directly determines the effect of figure to figure conversion.

If you want to fix the lines, then the noise reduction should be reduced to 0.2.

Curbing style pollution

Introduces how to generate images in WebUi with the subject as the focus, rather than the style of the drawing.9

To get the desired result, we change style to [style:10]

In Automatic 1111's WebUi prompt method, [x:10] tells Stable Diffusion until the tenth step before rendering the prompt.

By the time 10 steps are reached, a basic image has been formed, so styling pollutes the main component recklessly, instead acting like an IMG2IMG (which you can think of as a form of Img2img) over the existing image

You can experiment with this number depending on how many overall steps you've done. If you change the format to [x:0.2] (i.e. a number below 1) then it becomes a percentage, [x:0.6] tells SD to wait until 60% of the total number of steps before implementing x, and so on.

Size selection

It should not be linked to picture quality; size affects the subject to some extent, as it potentially represents the category chosen (e.g. vertical figures, horizontal landscapes, small resolution emojis mostly).

Picture quality can be manipulated using the Super Score guide.

Painting soldering!

With AiPhotoShop - the online tool for infinite extents of canvas you can "weld" your work continuously, very conveniently. Just add the --api parameter to the start command and open the web page to use it.

Image extents for some anime models are not supported because they are not Inpaint models (Sd has a dedicated Inpaint model.)


Trivial details

  • Emoticons

emoji(💰💶💷💴💵,🎊🪅🪄🎀,👩🚀) Emoticons are also available and very accurate, and Ai can recognize emoji after I tested it many times.

emoji performs well in terms of semantic accuracy because it has only one character.

emoji reference

  • Anime style

anime coloring will have a good effect, like anime screenshots. (If you encounter black and white, you need to specify the color of the character parts)


Optimize the effect of your work

  • Take a look at the goal

Think about what you want your work to be tagged with on the internet.

A sticker wouldn't say masterpiece, best quility, would it?

Another thing is that if the request is for a pixel work, you should remove some negative hints of conflict (if any).

  • Adjustment of order

The words in front anchor the color distribution of the image. What is important what goes first.

  • Semantic offset prevention

To prevent semantic shifts, prioritize emoji, and then use less complex syntax like with unnecessarily.

  • Keying?

The rest of the background of a buttoned image will be treated as black.

If you need a background, the effect is generally not very good, and it is recommended to fill in the color before operating.

  • Protect images from being cut by split lines

Negative tips to join

comic 2koma 3koma 4koma collage

It should improve.


If you want to reproduce on generation, you need the same seed, cue words, and other parameters (even size). But if the other side uses stylization, then you can hardly reproduce it.

-- Self-deceptive reproduction: figure to figure, noise reduction pulled to below 0.3... --


The core logic of cross-domain terminology is to narrow the scope of the specified data and start with platforms, domains, events outside of the screen content to enhance the effect.

Yes! You can use Film Glossary FILM GLOSSARY Photography Terminology in the prompt /jbeatty/COURSES/glossary.htm), Cinematic techniques, and painting terms (types) to control the basics.

Crossover! You can even use various thrill-sport terms to produce some special effects .... such as airborne

for example, depth of field, aperture, composition, camera position, elements of movement, a Chinese introduction to the glossary of artistic photography

BUT this effect may have the added effect of introducing data into the picture that you don't want to see in the style (e.g. live action rather than but secondary). Get the measurements right. Add Step and style cues as appropriate to improve.

You can also use platform names to limit the scope of the dataset, for example, terms like pixiv.

extended reading

Recommended use of terms contained in Danbooru

Danbooru tag group

Useful film terms

Types of shots

TV terms

Types of photography

Photographic terms

Extreme sports


Art of composition

Designing with Ai

  • Backgrounds

Once you get a slightly more saturated photo, you can process it in AI and use it as a background.

Use anime coloring to enhance the anime style.

  • Icons

Look at this

  • Emotion packs

The emojis work really well. The sticker with the western face script is very well done!

The documentation writer has tried the following prompt and it produces high quality emojis.

Don't ask for quality prompt words for stickers.

positive vocabulary

white background,
masterpiece,best quality may or may not be added, if quality is not enough

lowres can be added or not, add if not vintage enough

Chibi controls the character to be Q (big head, small body), add if not Q enough

sticker is to specify the style as a social media sticker, required

negative vocabulary

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

low quality can be added or not, if you want to be more flat, please add 3D to negative hints.

Size required square.

  • Character design



SEE HERE FromOrigin wiki

NAI in use out parameters

  • Use the full model

  • CLIP layer = 2

  • Use ema weight loading, set use_ema to true in the yaml configuration

  • reset sigma noise/strength to the default value of 1

  • set eta noise seed delta to 31337 (so that sigma noise/strength does not need to use 0.69 / 0.67)

  • If prompt has weights, convert the weights (WebUi to 1.1, NAI to 1.05)

  • Use the -no-half parameter to start the program (secondary)

NAI default model settings

steps": 28, "sampler": "[sampler]", "seed": [seed], "strength": 0.69, "noise": 0.667, "scale": 11.0,

Strength , noise 是 eta 和 sigma

scale 就是 CFG scale

NAI default SFW negative prompt

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


Add masterpiece, best quality in front of all prompt words

Clip skip 0, everything else is fine (afaik don't use supernet, v2, yaml, VAE)

Conversions - NAI and WebUi(SD) have different enhanced syntax

Prompts parameter bracket conversion

Convert enhanced parameters between NAI and Webui, related tgbot services M2NM2NBot

Related WebJS

Weight enhancement markers: NAI is {} and WebUi(SD) is ()

Prompt Magic Book

The Prompt Magic Book provides a library of ready-to-use templates (analogous to the Composition Book) with a convenient debugging log for Tag.

Prompt Magic Book Volume 1

Prompt The Book of Magic Volume 1.5

Prompt 魔法書製作委員會

Good parameters (style tends to be illustrative) 1

{an extremely delicate and beautiful}

Full classification of drawing mediums

List of anime characters/artists/stylizations/Pt files

Stylization: Doll Class

Stylization, Japanese Wiki

Stylization: 32 types

Artist List/SD1.4

Artist List/SD1.4/1,833 artists

Artist Museum

Stable Diffusion Artist Style Studies


sketch 可以让图片看起来像随手画的草稿
{{lineart}} 可以让线条变得很粗
{{{posing sketch}}}, {{monochrome}} 黑白草图
{rough sketch} 上了颜色的草图
monochrome+lineart 情况下一般只会让眼睛上色,强调发色后头发也可以上色
{{{monochrome}}}, {{{gray scale}}}, {{{pencil sketch lines}}} 做出的铅笔速写的感觉

利用 sketch,pastel color,lineart 的 tag 模拟一张图的绘画过程


chibi 可以画出低头身比的效果(二头身,三头身)
{{watercolor pencil}} 可以生成彩铅画
{{faux traditional media}} 可以做出签绘的风格
anime screeshot, 可以让画面变成动画风格
{{{retro artstyle}}} 赛璐璐风
{photorealistic}, {painting}, {realistic}, {sketch}, 厚涂
pastel color 和 sketch 搭配会有速涂的质感

杂志/设定集 风格

official art 变得更加官方一点
three views from front, back and side 和 costume setup materials 可以用来生成设定图
multiple views 会出现类似设定图
{character sheet} 会出现设定图
magazine cover 会把背景换成杂志封面,配合 office art 更像真实杂志(虽然字没法看)
magazine scan 类似杂志内页的风格
posing 会强调有一个动作,不至于出现混乱的动作(露出有六个手指头的手)
caustics 画面向主题聚焦,类似海报


人物数量 描述
数量 , one boy , one girl , two boy ,two girl,one_boy_one_girl(这是错误的)
人物画风 描述
质量提升参数 , masterpiece, best quality
原神 , Genshin Impact
萝莉 , female child , loli 画风差
人物样貌 描述
头发 hair
长发 longhair
短发 shorthair
眼睛 eyes
渐变颜色长发 gradient pink longhair
渐变颜色眼睛 gradient pink eyes
粗眉毛 thick eyebrows
猫尾巴 cat tail
猫耳朵 cat ears
动物耳朵 animal ears
毛茸茸的动物耳朵 animal ear fluff
刘海 bangs
两眼之间的头发 hair between eyes
眉毛后面的头发 eyebrows behind hair
锁骨 collarbone
斗篷(要在很前面才有效) cape
乳房尺寸 small breasts
出汗 sweating
颜色丝袜(和长丝袜冲突) white stockings , black stockings
长丝袜 thighhighs
女仆 maid
发带 ribbon
爱心眼 heart-shaped pupils
御姐/JK/辣妹? gyaru
肌肉发达 muscular
天使翅膀(要是形容人的第一个才正常) angel wings
颜色内裤(赠内衣) pink underpants
肚脐 navel
颈部颜色项圈 white collar
黑色皮肤 dark skin
撕裂的衣服 torn clothes
撕裂的裤子 torn legwear
开襟夹克(配合叉开腿特色) open jacket
异色瞳 heterochromia_blue_red
吊袜带(会和内衣冲突) garter straps
靴子 boots
眼罩 blindfold
流泪 tears
项链 necklace
眼镜 glasses
比基尼 bikini
湿衣服 wet clothes
透明衣物 transparent raincoat , transparent jacket , transparent tshirt
唾液(自动伸舌头) saliva
流口水(和唾液冲突) drooling
水手服 sailor dress
环境样式 描述
在床上 on bed
光线反射 reflection light
赛博朋克 cyberpunk, city, kowloon, rain
在地毯上 on carpet
在瑜伽垫上(它分不清什么是瑜伽垫,只知道色块比较大,所以要配合 one girl 用) on_yoga_mats
人物视角 描述
正面视角 from viewer
从上到下视角 from below
全身 full body
人物状态 描述
叉开腿 spread leg
露出腋下 armpits
举起手 hands up , arms up
爪子手 paw pose
站立 standing
行走 walking
吐舌头 tongue out
抬起腿 legs up
手放背后 arms behind back , hidden hands
衬衫 shirt
长袖 long sleeves
连帽衫 hoodie
褶边 frills
喇叭裤 bloomers
白色连衣裙 white dress
捆绑 bondage , bondage body , bondage foot , bondage hand
蹲下 crouch , squatting
真画风 photorealistic
跪下 kneel down
湿身 wet body

Market Adoption Survey

Here is the application of the stable diffusion (non-NAI model). Check out related discussions.



Album artwork

Allotment chart


  • There are many video platforms


  • Certain companies repackage open source projects as creator communities and charge users a subscription fee. For example, XX Gallery, XX Edition. There are more small programs.

  • Tencent QQ Small World allows users to use Ai for painting

NFT - ... ...

Last update: February 14, 2023 15:46:49
Created: November 2, 2022 07:44:44