This article has a dedicated section on large-scale model-related content. Before reading it, you can pay attention to it:
stable diffusion practical operation

Model download website
Domestic: https://www.liblibai.com
Foreign: https://civitai.com (Science Internet)

1. Development History

1 Official Mockup

Large model It takes a very long time to train a large model from scratch, and the cost is very high, millions of dollars:

2 CHECKPOINT

Retraining on the basis of the official large model is the CHECKPOINT model we saw, but the hardware requirements are still very high, and it is not very flexible to use

3 EMBEDDING

Later, the embedding model was developed, the file is small, the training cost is low, but the effect is not satisfactory, the advantage is flexibility

4 Lora and Hypernetworks

2. What is a large model

We can understand it as a packaged collection of various painting styles. Unlike midjourney, just write in the keywords what style you want, such as Chinese style ink, Disney, etc. As follows:

Stable Diffusion requires a large model to specify the style, such as the national style model, Disney model, etc. For details, refer to checkpoint. After selecting this model, all your keywords are determined according to this model, that is, Different models have different keywords, which is why SD is more difficult to learn.

3 Precautions for use

Through the second, we know that the keywords of each large model are different, so when downloading a large model, you must read the author's description and the author's keywords in detail, and learn through keywords, so that you can use this large model well.

text

serial number	model name	Introduction
1	Stable Diffusion commonly used official model V1.5	official
2	Stable Diffusion XL1.0	official
3	Anything series	Two-dimensional
4	Realistic Vision V2.0	realistic model
5	DreamShaper	portrait
6	MeinaMix	cartoon
7	ChilloutMix	Reality
8	majicMix series
9	AbyssOrangeMix2_hard.safetensors	comics

1. Official model

The bottom model/big model is the basic model that SD can draw. After installing the SD software, it must be used with the base model. Different basic models will focus on their painting styles and areas of expertise. Common modes, suffix ckpt, safetensors, size between 2-8G.
[1] Stable Diffusion commonly used official model V1.5 version, v1-5-pruned.ckpt.
Download link: https://openai.wiki/stable-diffusion-model-v1-5.html
[2] Upgrading to the new version [Stable Diffusion XL1.0 official version]
Download link: https://www.liblibai. com/modelinfo/77a9ca05a2c7c93ceb05f4dceae9ed97
【3】Anything series models, biased towards the two-dimensional
anything-v5-PrtRE.safetensors model can be used to generate various types of images, including portraits, landscapes, animals, cartoons, science fiction, etc. It has strong adaptability and versatility, and the effect shows a strong sense of reality and the ability to restore details.
insert image description here

Detailed introduction link: https://zhuanlan.zhihu.com/p/640797704

2. Realistic Vision V2.0, a large realistic model,

Link: https://pan.baidu.com/s/1wat_55oIRa-fSSleqpE1kw Extraction code: 5u8w
insert image description here

3、DreamShaper

This model works especially well on portrait prompts, creating a stunning piece of art that captures the feel and style elements of the subject. When it comes to landscapes, DreamShaper is able to create an interesting and complex image, with vibrant colors and captivating detail. Illustrations showcase vibrant colors and a variety of shapes that bring images to life.
Link: https://pan.baidu.com/s/1zZeTk3J4odbWqjwJle_1Bg Extraction code: ba47
This model is recommended for anyone who wants to create illustrations with their own AI models. Additionally, you can play with various settings to ensure the result looks more like a digital work of art. We recommend experimenting with different step counts and hinting engineering techniques with this model to get the best results.

4、MeinaMix

MeinaMix is one of the most popular models, and for good reason. This model is capable of producing stunning illustrations with a unique painterly quality. Much like the Anything model, MeinaMix focuses on the anime style when it comes to characters, but it is also capable of producing complex illustrations.
Link: https://pan.baidu.com/s/1ECn3AAU6CBj_6rOriFn7ig Extraction code: u24m
insert image description here

5. To download the checkpoint model, on station C or liblib, the label is checkpoint.

insert image description here

6、ChilloutMix-KoreanDollLikeness

When it comes to SD models, Chillout is essential. Even though many models that can replace Chillout have emerged on the market because of the stop of subsequent updates, Chillout is still a monument: the first real-life model that exploded.
Chillout has a dominant share in live-action drawing. It is a very good model, but maybe because it matured too early, it stopped updating soon. The latest version is the February 2023 version.

The model is divided into two precisions of 32 and 16, of which 16 precision can greatly save video memory, but for metaphysical reasons, I still use 32 precision.
Introduction and download link: https://zhuanlan.zhihu.com/p/632997666
insert image description here

7. majicMix series

Majic is also a model that can be called a god. The author is Mai Ju, a Chinese bigwig.

The light and shadow control is very good. This ability to control the shadow effect is a relatively weak part of SD compared to MJ before, but with majic, it makes up for it. Especially the eyes and demeanor of women are like wearing colored contact lenses. Even small freckles and moles on the characters can appear. What is especially shocking is the expressiveness towards men. From the muscles to the blood vessels, it is very real.

There is an anecdote, that is, after the AI bloggers Nausjia and Nana sauce refined their own lora and integrated it into sombra, the author directly integrated Nana sauce into the official version in the latest V5 version, which belongs to the real game Bo Yong was born.

The model recommends using Euler as the sampler.

And the author emphasized N times...don't open facial repair!
There are several majic series, namely:
insert image description here

7.1 majicMIX realistic: It focuses on the real system and is a general version.

Download introduction link: https://www.liblibai.com/modelinfo/bced6d7ec1460ac7b923fc5bc95c4540