There are many local installation tutorials for SD, so I won’t repeat them here. This is mainly to record the problems I encountered in the process of using SD Webui, and summarize some experience in improving the efficiency of drawing and the probability of producing good pictures.
Take a few pictures first to see the effect
Two-dimensional sister + Gundam = ?
Ok, the above is just a small part of the finished product, which is a work of attracting jade.
The effect of the picture produced by the parameters I set at that time was very good, and the fingers were basically intact. Let me tell you how to get started quickly.
Use of WebUI
Sinicization of webui
extentions, available, select the configuration address (default https://raw.githubusercontent.com/AUTOMATIC1111/stable-diffusion-webui-extensions/master/index.json ), uncheck the localization below, click load load from
Enter in the search box: CN, select Simplified Chinese localization, recommend using with Bilingual Localization.
After installation, click apply and restart (not Chinese yet)
Continue to select setting->interface, scroll to the end and select Chinese, and click save setting on the current page.
There are many official functional components in extentions, which can be downloaded by yourself.
model used
The commonly used Asian portrait chilloutmix series can be used to get started first, and then downloaded and placed in models/Stable-diffusion (usually around 4G to 8G)
Stylized LORA, after downloading, put it in the models/Lora directory, and you can produce pictures without using it (for example, the GundamRX782OutfitStyle_v10 I use here is generally around 140M)
parameter settings
1. Hires.fix fixes and enlarges
When testing the prompt and other parameters at the beginning, you can uncheck it first to quickly verify and output the graph. After confirming the other parameters, select them, and the hires steps generally get about 20. The more the steps, the better the quality, and the slower the speed.
If it is too low, it may cause gray edges and light colors in the output image.
It is good to use the default latent effect of the enlargement algorithm.
upscale by: 2 The default 2x zoom is enough.
Because of the function of AI zoom-in and repair, the picture with the resolution set to 512*768 is also very clear when outputting the picture.
2. Sampling method sampler : DPM++
Just start with those DPM++.
Sample steps sample steps : 20
The higher the value, the higher the quality, the slower the speed
3. Resolution
Generally, there are more 512*768 settings , depending on the specific model, the effect of random setting is very poor
4. batch count generates batches and batch size each batch quantity
batch count Execute generated batches
batch size The number of batches generated (consume more memory to improve the quality of generation)
5. Prompt words correlation CFG scale : 7
Generally, the default is good. If your prompt words are more specific (long and smelly), try adjusting to 8 or 9. The lower the value, the more creative, but the effect is hard to say.
6. seed random number seed: -1
The default is -1 at the beginning of drawing, or use the value set by other big guys. If it is a random picture -1, if you want to continue to produce what you want, just click it
Reuse the last random seed
SD is worthy of our further research, because it is open source, and the possibilities will be infinitely expanded.
As the saying goes, don't be afraid of floating clouds to cover your eyes, just because you are at the highest level
It is not easy to see this, and must be rewarded.
If you want to reproduce it, try it with chilloutmix_ni+ Gundam stylized lora.
The following resources include Gundam stylized lora files and some Gundam girls with good effects (some of them may not be published), as well as examples of prompts and parameter settings.