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

Perspective

Recommended use of terms contained in Danbooru

citing an image from the Japanese Wiki, author unknown

Noise

|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](https://github.com/xinntao/Real-ESRGAN) is recommended for noise reduction for anime illustrations.

rendering

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](https://heibox.uni-heidelberg.de/f/578df07c8fc04ffbadf3/?dl=1), file size is 1.9GB

[BSGRAN 4x](https://github.com/cszn/KAIR/releases/download/v1.0/BSRGAN.pth), file size 63.9M

[ESRGAN_4x](https://github.com/cszn/KAIR/releases/download/v1.0/ESRGAN.pth) with a file size of 63.8MB

[ScuNET GAN/PSNR](https://github.com/cszn/KAIR/releases/download/v1.0/scunet_color_real_gan.pth" to D:\stable-diffusio\models\ScuNET\ScuNET.pth), file size [SwinIR 4x](https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/003_realSR_BSRGAN_DFOWMFC_s64w8_SwinIR-L_x4_GAN.pth), 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)

Multi-figure/single-figure

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.

https://github.com/Mikubill/sd-webui-controlnet

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.)

GitHub

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)

some_art_styles_that_i_tried

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.

Reproduction

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... --

Associated

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

Sports_by_type

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

1girl,sticker,Chibi,Q,
white background,

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

BV14e4y1U7r9

Prompt

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

other

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

草图风格

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

艺术风格

杂志/设定集 风格

常用参数：SFW

Market Adoption Survey

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

3D

• On blender, Ai has rendering plugin

Design

• Microsoft 365 Tool Suite

• Generate portraits for the Age of Empires 3 Definitive edition of the game module

• Product and architecture design/sketch

• Painter cut and paste highlights

• Comic creation @852

• Novel illustration, AI painting background (I'm told the original model works well too)

Album artwork

• Generate image of singer with stable diffusion to put into video

Allotment chart

• Making weird diagrams as a memory tool

• Real use case of stable diffusion

• Application combination of stable diffusion

Self-Publishing

• There are many video platforms

Services

• 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 - ... ...