Hello, this time, I will explain some things you must know when using stable diffusion. But there are quite a few confusing parts when using it. Those of you who have used mid-Journey or dalle-3 may know to some extent, but there are quite a few things to choose from other than front. Checkpoint, Lora, Clipskip, and VAE. Those of you who are using it for the first time may be quite confused and the concept may not be well grasped. So, I will try to remove as much as possible of the content that contains difficult concepts and apply the words that we can use in everyday life. So, to understand stable diffusion easily, I brought some examples. If I explain it with an example like this, it can be understood much easier, so I brought an example. The tool called stable diffusion is a chef who makes the food we want to taste. You can think of a chef as a stable diffusion. Now, let me explain how stable diffusion works with this food called Tteokbokki. Tteokbokki is the image we want. To make these images, we need ingredients. Let's take a look at the ingredients. First of all, red pepper paste. Red pepper paste is the base. The base is the checkpoint. When making the base of Tteokbokki, it becomes a completely different food depending on whether you make it with black bean sauce or red pepper paste. Whether it's black bean sauce Tteokbokki or red pepper paste Tteokbokki. So, you can think of this checkpoint as a function that gives you this base feeling. Now, let's see how much difference there is. First, if you say, print out an image of a boy crying with a real-life checkpoint, it prints out the real-life image we always know. And when you give the same command in animation, the animation checkpoint gives you an animation-like feeling. In other words, like the base of Tteokbokki, when it's red pepper paste, it feels like red pepper paste, and when it's black bean sauce, it's like it's based on black bean sauce. The real-life checkpoint has a real-life feeling, and the animation checkpoint has a base of animation feeling. So, it would be better to first figure out what kind of image you want, and then use a base checkpoint that fits it. The second thing I'm going to explain is Lola. The concept of Lola is not that difficult. Lola is, in simple terms, a fish cake, cheese, dumpling, and rice cake that goes into Tteokbokki. You can think of it as playing that role. You can't change the taste of this base. But it's a feeling that can affect you to some extent. Let me give you an example. When I applied Lola to the same real-life checkpoint, I applied animation Lola. Now, when I did that, a boy was smiling, but it's a little awkward. If you mix animation with a real-life checkpoint, you can get a little awkward feeling. But if you mix animation Lola with animation checkpoint, you can get a much more natural animation form. So, Lola affects the checkpoint to some extent, but it can't completely change the taste. It's like a fish cake or a dumpling. The third thing I'm going to explain is VAE. VAE is like a seasoning. We made Tteokbokki like this. Some people's taste is a little bland, and some people's taste is bland. In that case, if you add ramen soup or seasoning called magic soup, it can be a Tteokbokki that can be a little balanced for everyone's taste. You can think of it as VAE playing that role. To give you a little more explanation, you can think of it as a fix. When there is no VAE, if you look at the left picture, it feels a little dull and damp. But if you look at the image of VAE applied, it feels a little clearer and cleaner. You can think of it as a slightly low filter. For those who are using it for the first time, we use this 840,000 VAE a lot. Later, there may be a little relationship between VAE and Lola and the checkpoint, but at first it's not very important. So, it would be easy to think that VAE plays a role of MSG. The last thing I'm going to explain is Clip Skip. Clip Skip is like this. Table Diffusion is a chef. You can think of it as the chef's recipe-thief ability. It's simple. Usually, Clip Skip is set to 1. It's set to 1 at first. You can give it from 1 to 12. The higher the number, the more ability to understand the prompt we use. But usually, when you learn the checkpoint, Clip Skip is used. Because it gives you this level of value, If you use Clip Skip at this level, You can get a better image with a higher probability. You can think of it that way. Now, when we ask for it, For example, if you listen to Tteokbokki, Boil it for 3-4 minutes, Put the finishing in a bowl and bring it to me. I asked you to make it. Now, if Clip Skip is low and the understanding is a little low, AI says, Oh, so just stir-fry it, stir-fry the ingredients, You can put it in a bowl. You can make a messy image like this. But if you set Clip Skip to the correct value, It's a little cleaner and more sensible. You can make an image. So I explained it with Tteokbokki today. Tteokbokki is also a mixture of all ingredients and the chef's ability. Like delicious food coming out, This stable diffusion AI image is also a checkpoint. And Lora, VA, Clip Skip, These basic things have to be mixed very well. I think you can get a good quality image. Yes, I've explained the essential things you need to know today. I hope it helped. I'll see you next time with a little more informative content. Thank you.
