The expression of the face is not perfect, but it is a little more upscale than the left side. You can see that the facial expression has improved a little more. You can also see that the details around this are a little higher on the right. Hello, nice to meet you. This time, I will introduce the function called Hi-Res Fix in Stable Diffusion. The version I'm using now is 1.5.2 version. If you use the 1.6.0 version or 1.7.0, it has been updated recently. They may use a little differently, but the function itself is similar, so check it out. When I have some time, I will explain how to apply it in the Hi-Res Fix 1.6.0 version or the latest version. Today, I will introduce how to use Hi-Res Fix in the 1.5.2 version and what functions it has. Hi-Res Fix is usually used to upscale the image after creating it in the Text-to-Image or to improve the image quality. So let me explain how to use it. If you look at the Text-to-Image, there is a checkbox called Hi-Res.Fix under the Sampling Method in the 1.5.2 version. If you check this checkbox, other settings will open like this. Then, when you create an image in this Text-to-Image, a new image is created through this Hi-Res Fix function. There are thousands of ways to use Hi-Res Fix, but it is usually divided into two. The first way is to fix Hi-Res Fix when you just create an image and create it from the beginning. Or, if you create an image in the Text-to-Image, and if the image is good, you can use it in the way of upscaling with the seed value of the image. You can see that it is divided into two categories. Of course, there are many people who use it in other ways, but you can think of it as using this function by dividing it into two categories. Then, I will explain a little about the basic functions of Hi-Res Fix. Sampling Method and Sampling Steps are included, so I will exclude the explanation. If you look at it, Resize from 512x512 to 1024x1024 means that the existing image is 512x510, but it has doubled now. It means that I will upscale it by twice. So where do you adjust this value? If you come down below, there is an upscale by. This is the value that adjusts the size. If you type 2, it will be doubled, and if you type 3, it will be tripled. But I usually don't apply more than twice when using Hi-Res Fix. Because there is a specification problem with the local PC GPU, and I don't do that much with Hi-Res Fix, so I usually give it less than twice. So I use it about 1.5 times. And if the limit of the GPU capacity is more than 1024 size, or more than 1000 pixels, I proceed by considerably lowering the upscale. Of course, this is different for each person, but if you have a large amount of graphics card, you can do more than twice, so you can use this to your liking. And since it supports even the decimal point, you can enter the decimal point and adjust the size. This was the upscale by. Next, there is an upscaler. This is how you're going to upscale. There are many ways. You can apply this by using it yourself. When you upscale an image that feels like a real-life, R, E, S, R, GAN, 4, X, PLUS. When this feels like a real-life, apply this upscaler. When it feels like an animation, apply the PLUS ANIME 6P below it. I usually use this a lot. I think these two are the best, so I choose one of these two models in the basic upscaler. Next, there are Hi-Res Step and Denoising Strength. I explained Denoising Strength in the previous Image to Image. First, let me explain Hi-Res Step. I think Hi-Res Step is almost the same as Sampling Step. I don't think you have to pay this price. Of course, there are many opinions, but I've used it myself. I didn't feel that big of a difference. Even if you give 0, the quality and image size or image quality I don't think it's that big of an impact. Or even if you raise the value, it affects the speed of image production. I don't think the quality has changed that much. So this is usually 5 to 14 or 15. But it doesn't matter if you don't touch it. I think you can adjust it to your taste while using it. Finally, Denoising Strength. It's the Denoising Strength I've always emphasized in Image to Image. It's similar here. When I applied Denoising Strength, When it's above 0.5, it changes the feeling of the existing image. The image is upscaled and when it's below 0.5, it's upscaled while keeping the existing feeling. I'll explain this again while I'm creating it. So when Denoising Strength comes out, if you want to keep the feeling, If you want to try a completely new attempt below 0.5, I think you can give me more than 0.5. This is my personal opinion. If you have a better opinion about Denoising Strength, I'd appreciate it if you could let me know in the comments or separately. Now, let's get started with Hi-Res Fix. First, turn off Hi-Res Fix and create an image. I'll show you how to upscale with the image. First, turn off Hi-Res Fix like this. Enter negative in the prompt window. I'll make it simple. A boy is running. I'll create it with DPM++ SD Keras and 7.30. Yes, the image is created. Let's check it out. Of course, I didn't apply AD Tailor for the details. I won't look at the details. Let's see how the image changes when Hi-Res Fix is applied. With this created image, To apply Hi-Res Fix in the Text2 image, You have to apply a seed value. Seed value is a concept that we use in stable diffusion or in the production AI tool, DALI or MidZone. When you create an image, the tail always follows. To put it simply, Seed value is the ID value of the image created like this. So if you apply Hi-Res Fix with this ID value, You can apply the seed value you just made to upscale a similar image. Now, let me show you. First, you need to check the seed value of the image. There are two ways to check the seed value of this image. First, if you scroll down, there are letters written all the way down here. If you read the letters, there's a seed and a number down here. There's a number, and if you copy this number, It's going to be a seed-1. If you paste it here, it means that you will continue to create images with this number. So this similar image keeps being created. So the second way is, Oh, then every time I find the seed value, Do I have to enter Ctrl-C, Ctrl-V like that? No. There's a simpler way. If you look to the right, there are two buttons. One is a dice, and the other is a recycling button. If you press the dice button, the current fixed seed value is solved to minus 1. So if you press the dice button, You can think of it as a random seed value. If you press the recycle button on the right, The seed value of the image I just created. You can see that it's applied like this. So, if you use the recycle button and the dice button well, you can save more time. Now, let's apply Hi-Res Fix to this image to proceed with upscaling. Now, press the recycle button here, You can leave the rest of the values as they are. Click the Hi-Res Fix button here. If you click it, you'll see the set values. Upscaler feels like this image is real, So I'm going to change it to Hi-Res Fix. And as I explained in the Hi-Res Step, I don't think I need to give it a value of 4. I don't want to change much in this image. I'll give you about 0.35. Then, leave the horizontal and vertical values as they are. I'll give you 0.35. And then, I'll give you 0.35. I told you that I'm going to adjust the number. So I'm going to give you a value of 1.3. And the rest of the values are the same. After giving the batch count to 1, I'll create it at the end. Then the image is created like this, I can see the silhouette little by little. How it's made, I'll give you a value of 1.35. And then, I'll give you a value of 1.35. I can see the silhouette little by little. How it's made, The image is all created. Let's compare the two images and see how they changed. First of all, we need to check that the top size is 665. So, we need to check if the black has grown to this size from 512x512. As I said, I gave you a weak denoising stress. There should be no big difference from the previous image. At the same time, the pixels should be bigger while maintaining the resolution. Let's take a look at that. Let's go to the output folder and look at the size first. The PNG file that the left image first created. You can see the size of the photo 512x512 here. And the second file that applied the high-res fix that was created, It says 664x664. Yes, if you create it like this, there are times when there is a difference of 1 pixel. Please check that. Now, you can see that it's upscaled like this. I'll turn on two and compare them side by side. Let's compare the two. The text 2 image that the left image first created. The image on the right applied the high-res fix. And no other editing device was applied. Now, you can see that the right image has a little more quality than the left. The expression of the face or the right is not perfect, It's a little more upscaled than the left, You can see that the facial expression has improved a little bit. You can see that the details around this are also a little higher on the right. However, this is also a concept that creates a new image. I gave you a low denoising value earlier, but you can see that it has changed little by little. The overall feeling of the character is the same, If you look at this costume, you can see that it is buried in blue paint. You can see that the hand has changed a little bit. So when you apply the high-res fix, If you want to increase the size of a 100% perfect result, You don't have to apply this high-res fix, You just have to increase the size in the extras. But the reason I use high-res fix is As I just compared while raising this image, It's a little bit of a detail. So when I write the high-res fix in the text 2 image, When I make a new text 2 image, Many people use it as if they were fixing the high-res fix. Now, I explained two examples at first. After checking the image once like this, Those who use high-res fix in the way of upscaling, Or to avoid this low-quality image from the beginning, I said there are people who use it with the high-res fix fixed. I'll make it like that. Now, this time, the seed value is fixed now, If you make it in this state, you will get the same image again. Or a similar image will come out. Now, then, I'll solve the seed value with the syringe, I'll give you the same upscaler, and I'll give you the same denoising 0.35. Now, and I'll make it in this state without changing the text 2 image and the prompt. Let's check how the light comes out. Now, the image is created. Shall we check it out? Now, if you look at it, you can see that it's very different from what you felt when you didn't apply the high-res fix. When we just made it 512x512 without applying any other correction device, The detail is quite low. But in the first place, if you apply high-res fix like this, If you set up the denoising in advance and increase the size by a few times with the upscale, You can create a pretty good image from the beginning. But that doesn't mean you can get a 100% good feeling. So, if you've tried it yourself several times, I make it like this in this case, I saw the image was selected first, If you say that you should use a good seed value there to upscale with high-res fix, I think it would be good to make it using the method I first told you. Today, I briefly explained what high-res fix is. In the end, if you want to get a sense of this high-res fix, you have to deal with it a lot. If you have something that doesn't work well and is blocked, If you come to the comment or cafe and leave a comment, I'm answering when I can check. I think it would be good to use these channels a lot. That's it for today's lecture. Next time, I'll find more useful content and explain it to you. Thank you. Thank you for watching.
