Install the Stable Diffusion webui

Renderings:

1. Preparations

Before downloading AIthe painting tool, you need to download some running environments on the computer in advance

1. Download python, git

First of all, the local machine is preferably Nvidia's N card, and it needs at least 4GBvideo memory to run locally. Of course, A card can also be played, because drawing has high requirements on the graphics card. Otherwise, even if the installation is successful, the resulting image may be completely black or cannot be drawn at all.

When installing, you need to turn on Magic Internet, otherwise you can't download some dependencies needed for painting

In addition, a local pythonenvironment , and the official Python 3.10.6 version is required to be downloaded

Among pythonthem the download address: https://www.python.org/downloads/

In addition, installation is required Git, download address: https://git-scm.com/downloads/

pythonAfter the installation is complete, we run the following command to pipupgrade to the latest

python -m pip install --upgrade pip

2. Download the Stable Diffusion webui

The first uses Gitthe command :

git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui.git

The second is to directly download the compressed package file, as shown in the figure below:

3. Run start

After downloading to the local, if you are windowsthe system , double-click directly in the folderwebui-user.bat

If you are MACOSor Linuxsystem, run the command./webui.sh

It will take a long time to download various dependencies when starting for the first time, and the latest version has added a 4GBnear default model v1-5-pruned-emaonly.safetensors, which also needs to be downloaded, so just wait patiently.

If you want to speed up the download speed, you can download the model in advance, put it in models/Stable-diffusionthe folder , and then start it again webui-user.bat, so that you will not download the model again

The default official download address of the model: https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main

If you feel that the official default download is slow, I also posted the download address of Baidu Cloud’s default model in the comment area, please choose by yourself

If the installation process goes well, open it directly in the browserhttp://127.0.0.1:7860

A successful visit means that the installation is successful, and the opened page is roughly as follows

2. Download the plug-in and localize it

Chinese installation

Way 1: Install from available

Click Extensionsthe button , then click Availablethe available extensions button, and finally click Load fromthe button , it will be able to retrieve all available extensions from the index link behind for your own download

When the list of extensions is loaded, we search for the Bilingual Localizationlanguage localization plug-in, and click to installdownload it.

Way 2: Install from the website

Click first Extensions, then click the third button belowInstall from URL

You need to manually fill in the Chinese link address for installation. Here we can githubchoose the link or go to giteethe link above in China.

The open source warehouse of the Sinicization plug-in is sd-webui-bilingual-localizationOr stable-diffusion-webui-chinese, search for it, then paste the link in, and Installclick

https://github.com/VinsonLaro/stable-diffusion-webui-chinese.git

Way 3: Install from the website

We directly find stable-diffusion-webui/extensionsthe extension directory, then open a command line window and run the following command in that directory

git clone 仓库地址

Enable Sinicization

Method 1: Click Settingthe button , then click Bilingual Localizationthe column on the left, and Localization file select the language in the drop-down box

Method 2: Click Settingthe button , then click User interfacethe bar on the left, and scroll down to the bottom to see the opening window Localizationgof

After choosing, we need to click Apply settingto save the configuration, and then click to Reload UIreload the interface to play happily

Bilingual Sinicization

Bilingual Chinese warehouse link: https://github.com/journey-ad/sd-webui-bilingual-localization.git
The simple Sinicization above is just that the interface is fully converted to Chinese. If you want to display both Chinese and English, you need to use a bilingual plug-in
It is also through any of the above three methods to download the Chinese plug-in, you need to enable it to see the effect
At this time, we need Settingto set User interfacethe under the column on the left side Localizationgof None, and Bilingual Localizationthen Localization file select the language in the drop-down box of the column, and UIthen restart

Use

Renderings:

Interface parameters

Before using Stable Diffusionopen source AIpainting, you need to understand the painting interface and the meaning of some parameters

At present, the effect of AI painting on Chinese prompt words is not good, try to use English prompt words

The most important parameters are as follows:

Prompt: Positive prompt words, that is tag, the more prompt words, AIthe more accurate the drawing result will be

Negative prompt: Reverse prompt, that is, reversetag

Width / Height : The image size to be generated. The larger the size, the more performance and time-consuming it takes.

CFG scale: the degree ofAI preference for the description parameters . PromptThe smaller the value, the more the generated picture deviates from your description, but the more logical; the larger the value, the more consistent the generated picture is with your description, but it may not be logical.

Sampling method: Sampling method. There are many kinds, but there are only differences in the sampling algorithm, there is no good or bad, just choose the one that suits you.

Sampling steps: Sampling step size. If it is too small, the randomness of sampling will be very high, if it is too large, the sampling efficiency will be very low, and the probability of rejection will be high

Seed: random number seed. A random seed when generating each image, this seed is used as the basis for determining the initial state of the diffusion. If you don't understand, just use random

Restore faces: face repair, optimize the face, the principle is to call a neural network model to repair the face and affect the face. Repair the face of the character in the picture, but if the face repair is turned on for a non-realistic character, the face may collapse

Tiling: Seamless texture is an old-fashioned optimization technology, that is, the matrix multiplication optimization CUDAof , which affects the drawing speed and reduces video memory consumption, but after actually selecting it, the normal picture may not even come out. can't beat

Highres.fix: High resolution fix, by default Vincent diagrams produce very chaotic images at high resolutions. If you use high-definition repair, it will first generate a picture according to the specified size, and then expand the resolution of the picture through the enlargement algorithm to achieve the effect of high-definition large picture

Sampler

Euler a: It is an illustration, tagthe utilization rate is second only to DPM2and DPM2 a, the ambient light effect is good, and the composition is sometimes weird

Euler: Soft, also suitable for illustrations, good environment detail and rendering, deep background blur

Heun: The average quality ratio of a single Eulerunearthed Euler ais higher than and , but the speed is the slowest, and the high stepperformance is good

DDIM: Suitable for wide images, low speed, high stepperformance is good, tagif the negative is not enough, it can play casually, the effect of ambient light and water vapor is good, and the realism is not good

DPM2: This sampling method taghas the 80%highest utilization rate of

DPM2 a: Almost the DPM2same as , but there may be close-ups when the characters are

PLMS: The quality of a single drawing is second only toHeun

LMS: Texture OA, saturation and contrast are low, more animation style

LMS Karras: It will be greatly changed to an optimized style, and the realism is not good

DPM fast: This is a test tool left by the developers of the upper world, not suitable for use by magicians

Vincent diagram (txt2img)

Vincent image (text2img): Generate corresponding images according to the description Promptof

Select 绘画模型（Stable Diffusion checkpoint）► 文生图（txt2img）► 填写正向提示词（Prompt）► 填写反向提示词（Negative prompt）► 选择采样器（一般选择DPM++ 2M Karras）► 设置采样步骤（一般设置30）► 设置长宽（Width / Height，推荐 1024 * 1536）►点击生成（Generate）

According to the above steps, the picture can be generated from the text. Generally, we only Promptneed to , Negative promptwhich is usually fixed, and limit the watermark, low quality and other possible logic errors.

Image generated image (img2img)

Image-generated image (img2img): generate another new image from an image based on the characteristics Promptdescribed

Select 绘画模型（Stable Diffusion checkpoint）► 图生图（img2img）► 填写正向提示词（Prompt）► 填写反向提示词（Negative prompt）► 选择采样器（一般选择DPM++ 2M Karras）► 设置采样步骤（一般设置30）► 设置长宽（Width / Height，推荐 1024 * 1536）►点击生成（Generate）

Upload a picture to Tushengtu and test it. Here we use the photo of Wenshengtu just now. According to the above steps, click Generate to see the effect

It can be seen that AIthe picture generated by the picture is actually based on the picture you uploaded and yours to Promptgenerate a photo similar to the one you uploaded

label weight

When entering Prompttags , sometimes you will write whatever you think of, and the weight written in the front will be the highest by default

stable DiffusionYou can also increase the weight of keywords through specific symbols in

where ()is to strengthen the weight, []is to reduce the weight, is{} the "implementation" of for embeddingNAI()<>

For example: (eye) > [eye], (eye:1.4)will add approx. more than normal 40%, (eye:0.6)will decrease approx.40 %

Model download

Model website recommendation: https://civitai.com/

The drop-down box in the upper left corner can be used to select the model, where v1-5-pruned-emaonly.safetensorsis the default model

The picture drawn by the default model is ugly, so we need to download a better and more suitable model to draw

Each model focuses on different styles of painting. Some are suitable for drawing people, some are suitable for drawing animation, and some are suitable for drawing landscapes. Choose the model that suits your own painting style and download it according to your needs.

Download the model, put the model file into models/Stable-diffusionthe folder , reload UI, and you can see

The general difference between the models

name	need	Effect	Remark
stable-diffusion (4GB emaonly model)	2GB video memory to start	The drawing speed is 10s, and the maximum single output is 920×920	suitable for drawing
stable-diffusion (7GB full ema model)	4GB video memory to start	With the weight of the last training, so it costs video memory	suitable for training
waifu (Float 16 EMA model)	2GB video memory start	The performance is close to stable, but the video memory usage is slightly higher	suitable for drawing
waifu (Float 32 EMA model)	2GB video memory start	The performance is close to stable, but the video memory usage is slightly higher	Suitable for drawing, the picture quality is actually not much different from 16
waifu (Float 32 Full model)	4GB video memory starts	The performance is close to stable, but the video memory usage is slightly higher	Suitable for drawing or training
waifu (Float 32 Full + Optimizer model)	8GB video memory start	The performance is close to stable, but the video memory usage is slightly higher	suitable for training
Naifu (4GB pruned model)	Minimum 8GB VRAM & 8GB VRAM	closer to the official	suitable for drawing
Naifu (7GB latest model)	Minimum 8GB video memory (floating 10GB upwards)	closer to the official	Suitable for drawing or training

Note that the video memory here refers to the size of 512X512. Under the default configuration, the video memory that the software needs to occupy when outputting a picture. 2GB video memory starts, which means that the actual video memory of your computer graphics card must be at least 3GB (because the system desktop and browser display also need to occupy part of the video memory)
By adding various "optimization" parameters, you can exchange for a reduction in video memory usage through a partial decrease in performance.
Nafu model name description 1: animefull-final-pruned = full-latest = NAI full model (including NSFW)
Nafu model name description 2: animesfw-latest = NAI baseline model

References:

Stable DiffusionOpen source address: GitHub - AUTOMATIC1111/stable-diffusion-webui: Stable Diffusion web UI

Use of the latest version of Stable Diffusion open source AI painting tool

AI Tag Generator Geek Feitu AI Tag Generator

AI painting Stable Diffusion installation, tutorial integration package download