512x512 size and finally, if you compare the image that is twice upscale, you can see that there is a huge difference. You can see that the details have improved to an incomparable extent. Hello, nice to meet you. This time, I will introduce SD Upscale from Stable Diffusion. SD Upscale is a tool that allows you to change the image you have created from the image-to-image to a high-definition image in low-definition. It is an upscale method that can be used basically without installing a separate expansion function. There are various upscale methods, but I will explain how to use this SD Upscale that can be used basically among them. To use SD Upscale, you have to go to the image-to-image and upload the image and then work on it, but I don't have an image I've made separately, so I'll make it in text-to-image first, move that image to image-to-image, and see what changes there are through SD Upscale. First, if you are going to proceed in text-to-image like me, you can follow my method. And I'm doing it in 1.5.2 version, so please note that it can be a little different in the latest version or 1.6.0 version. But the framework is almost the same, so you can see and understand that it works with this principle. Let's make it under the same conditions to create an image. I will create it with the negative prompt and the simple positive prompt that I always use. I'll call the negative prompt from the file, and I'll say that the girl is smiling on a clear day in the park, and I'll apply DPM++ SD-Carras, and the step is 30, and the size is 512x512. I'll leave it as this condition and create it. Yes, the image is all created, and let's take a look. Now, if you look at it, the image is not very modified, but you can see that it came out pretty good. Now, since it is 512x512 size, if you enlarge this image, it will break very quickly. So, I'm going to upscale this size to twice the size, and I'm going to upscale it with almost no change in this image. It's similar to the Hi-Res Fix I explained before, but you can think of it as a little different. Upscaling is the same concept itself. Now, how do you upscale this in image to image? It's simple. The way to send this image I made to image to image is 1.5.2 version is convenient because it is written like this. There is a send to image to image. Please click this button. Click to automatically go to the image to image tab, and you can see that the photo has moved over here at the same time. Now, how do you do the basic upscale in this state? It's simple. Now, if you come down, it's almost the same condition because it's image to image. Now, this resize mode, sampling method, step, and size, strength, all of this goes in. So, I said that denoising strength is the most important thing in image to image. If you put a lot of denoising strength in this upscale, the original image will be damaged. So, I'll just give you 0.3 upscale. Now, you have to choose upscale, and if you look down at the part you choose, at the bottom, now I chose it earlier and it came out like this, but you have a script. At the bottom, it's probably scripted like this screen right now, and it's probably floating in non. Now, if you click this non, there are some functions, and there is a function called SD upscale here. You can click this. Click. If you click, tile overlap, and scale vector, and finally, upscaler. There's something like this, but you don't have to think too hard about it. How many times will the tile overlap overlap per pixel? Well, it's this number, but you can just leave this as the default value. And scale factor is a little important. Scale factor is how many times you're going to increase this image. So, if it's 2 now, it means that you're going to upscale this image by 2, and if you write 3, it means that you're going to upscale it by 3. You can understand that. And it's not just the numbers that are 1, 2, 3, 4, but 1.5, even the decimal point. You can use it according to the situation and the amount of GPU. Lastly, the upscaler, I've explained it once before in Hi-Res Fix. What's close to the real thing is RESRGAN4X plus is the best of the basic models, and in the animation feel, I explained that this plus anime 6P is a little better. It's the same. And in the case of this upscaler, I'll introduce another way to download it from another place later, and I'll explain it when I introduce another way to upscale. First, you just need to know the basics. It's an animation feel, so I'll choose 4X plus anime 6P and proceed. Click and think that all of these values are applied and made. Now, let's check it out for the last time, and click the Create button to check the results. Now, the image is complete. Let's check it out. Now, when I look at the image, I can feel that the details have come to life and become a little clearer. If you compare the left and right now, it feels similar, but anyone can see that it's a little fluctuated. Now, in the case of the right, I'll show you a separate comparison later, but you can feel that the details have come to life overall, and the face has changed a little. The reason for the slight change is that the denoising value was given 0.3, so there is no choice but to change to some extent. Still, you can see that the basic frame is maintained and the overall upscale has been done. The current image was an image that upscaled 1.5 times. Now, after upscaling it twice, the first text2 image, 512x512, and then 1.5x size, lastly, I'll check how it changes when I double the scale factor. Again, I'll give the same denoising strength to 0.3 and create it. Now, the image is all created, so let's compare the original 1.5x and 2x size and see what the difference is. First, let's compare the original image and the 1.5x upscale image. I didn't zoom in, but you can see that the basic size is already different. The left side is 512x512 size, and the right side is 1.5x upscale image. It's almost the same, but you can see that the details have changed a little. Now, when you enlarge the 512 image, there are a few dots and black colored areas under the eyes, and if you look at the 1.5x upscale part, you can see that this side is gone. And you can see that the overall texture and detail of the skin have been improved. Now, you can see that the details of the pupils have been improved and the enlargement has not been broken so much. Now, let's compare the 1.5x upscale and the 2x upscale. Now, I've definitely run the image, but you can see that the image size itself has changed completely. You can see that the overall feeling is the same, but I said that the little change is due to the denoising value. Let's compare. The left side is 1.5x upscale image and the right side is 2x upscale image. If you zoom in like this, you can see that the details are definitely better, but the disappointing part is that it changes slightly. So, since the face-like part is slightly modified, it is better to use this part. Then, finally, if you compare the original 512x512 size and the final 2x upscale image, you can see that there is a huge difference. Now, you can see that the details have improved to an unparalleled level. However, this SD upscale is the most basic upscale method. So, this upscale is not the best way. And the biggest drawback is that it follows the GPU performance too faithfully. So, if your GPU specifications are not good, it will be difficult to upscale more than twice. If you look down, the scale factor is 2x now, but this follows the GPU performance quite a bit. So, now it's 512x512, so it may be well upscaled without any errors on my PC. However, if you increase this scale factor 3 or 4 times, and if your GPU performance is really not that good, it may not work because of an error. I'll try to increase it by 6 times, 4 times, and 3 times and see what happens. If you adjust the factor 3 times and create it, it's 3 times now, but it didn't burst. But another drawback of this SD upscale is that it can burst at the last minute. So, pretend that the process is almost complete like this, and there are times when the white bar comes out at the end and doesn't work. And now it's upscaled 3 times, but if you upscale this image again, the result may be different. I'll try it twice here. If you do this, you can feel that the GPU is quite burdened from now on, and there are cases where errors occur or do not work. So, next time, I'll talk about how to upscale this upscale as much as I want in the next video. Yes, I upscaled it twice in this 1536x1536 size, and the GPU couldn't stand it and declared a stop. So, the stable diffusion has an error and a break. Like this, SD upscale is a good function when upscaling an image in a basic situation, but later we may want to upscale the image 2 times, 3 times, 4 times, 5 times, 8 times, but there is a limit to this function due to the capacity of the GPU. However, if you have a suitable GPU without this function limit, I will introduce an upscale method that can be done up to 8 times, 16 times next time. That's it for today's lecture. I'll prepare more useful content next time and see you. Thank you. Thank you for watching!
