Hello In this video, I will introduce how to make a game icon using Stable Diffusion, which has recently become a hot topic. Rather than dealing with Stable Diffusion in depth, this video briefly explains only the essential parts in terms of game icon production . You can watch it without any burden. When making a game, there are times when you have to draw so many icons. It would be best to just buy it somewhere, but it is not easy to actually do that because icons must reflect the game's content or design, and the art style must also be consistent. When an image-generating AI like Stable Diffusion appeared , the first thought that came to my mind was that it would be nice to create an icon here. After installing and running Stable Diffusion Web UI, you can see a screen like this in your browser. I won't explain it here because there so many videos about its install. If you look at the top left, there is a training data called Checkpoint that will be used to create an image. In this video , I will use Deliberate Checkpoint, which is one of the most used ones. Then, I will create an apple image for the apple icon. Write down the image you want to do. I will write the basic content of the apple and use it as an icon, so I will add keywords such as Still life, Painterly, Game Art, High Quality, Studio Lighting. and in the Negative Prompt just below , I added a couple of things. Because there are cases where the letters of the collected data are exposed in the process of creating an image. Then, I wrote Ground Reflection which makes the mask work difficult. Symmetry . Symmetry, I added it because it sometimes creates symmetrical images like decalcomani . The image itself seems to have come out well. However, if the background color is more consistent, the Mask work will be much easier, and if you can control the location, size, and timing here, it will look much more organized and stable when you see several icons together in a place like Inventory later. Then I will add Black Background to the prompt and create an image again. Black Background Black Backdrop in the prompt seems to be reflected to some extent, but as the floor or other elements intervene, it is virtually meaningless and overall The result was worse than the first time. I tried changing the prompt continuously, but the conditions mentioned above were not met. As I found out later, using an extension called ControlNet and Text2Image can also meet those conditions. However, here is a more convenient icon creation It's Image2Image creation. . In Image2Image, an image is transformed into another image according to prompt and other settings. Here, two important settings are CFG Scale and Denoising Strength. CFG Scale decides how much the prompt content has effect on the image creation. Denoising Strength determines how much the AI will change the image. I'll import the prompt I used earlier and create an image with default settings . I increased both values to reflect the content more and make the image more like an apple. An image of an apple was created with the background color, position, size, etc. controlled. I realized that it is very difficult to control color in Image2Image. So, from the beginning, I gave the starting image color and made an apple. I confirmed that I could create an apple image in this way, and then I made a tomato using the same red circle. The tomato was also made well, but There were two problems. One was that the tomato looked elongated up and down because it started from a circle, and the other was the location of the top . By changing the starting image like this, the tomato was able to maintain a certain point of view, just like the apple. There is still one problem left. Since I don't have enough experience in the prompt writing, the lack of information leads to the lack of image quality. I went back to the site where I downloaded the , looked at the works made with that checkpoint, picked out the ones I liked, and decided to copy some of the prompts. Name of Famous Artists, Trending on ArtStation and CGSociety, High Detail, Sharp Focus, Dramatic, Photorealistic Painting. It is not possible to explain exactly which one influences where and how in the final image, but as a result , I was able to see that the painterly feeling and detail were much more alive . I also made a bread icon in the same way. I tested the mask work myself in Photoshop to see if it was easy. I was able to remove the background and the border very easily. I started with simple forms such as apples and tomato bread, and since I tested natural or organic objects to some extent, I decided to make an icon with a more specific shape. So this time, instead of drawing a simple shape in Photoshop, I made an image by rendering in Blender . With 3D rendering, I can make an image in which lighting information is contained. I made this dark red box. What I intended is an old book covered with old leather. So, I added Old Leather Covered Book to the prompt I added more. Gold Ornaments, Dragon Engraved on the Front Cover. This time, the contents I wrote in the Prompt were not reflected well than I thought . I decided to put more information into the starting image by reviving my experience in Tomato. So I prepared this image. You can see my intention clearly by newly added things to the prompt. Magic, Spell, Rich Gold Ornament, Large Blue Gem, Lots of Leather Engraving is added. I created an image. The information I put in the starting image and the content are well reflected in the prompt . However, I couldn't see the splendor I expected. So I went down to the settings below and increased the value of Denoising Strength, which means the degree of image deformation by AI. I made a ring. I decided to make a shield at the end of the experiment. I couldn't create a dragon properly when I made a leather book earlier. This time, I decided to try again by putting a large lion's head at the center of the shield . Shield, Medieval, Fantasy, Warhammer, Large Gold Lion Head, at the Center has been added to thee prompt. Let's make an image. These images came out. There are places that reflect the intention well, and there are places that are not at all. I decided to create an image using the one closest to my intention as the starting image . When I made the magic book earlier, when I increased the Denosing Strength value in the settings, the result was better, so this time I set the value to 0.65. Let's do this now Among the images created in this experiment , I was able to obtain images that best reflected my concept. However, instead of this completeness, one thing was lost . It is slightly moved and rotated. You can see that the created image is facing the front. So I lowered the Denoising Strength to 5.5 and recreated the image. I was able to get this result. You can see that the detailed details have become much richer while maintaining the perspective. When the starting image is relatively low-res and the details were slightly crushed, it was close to a painterly feeling. On the other hand, the upscaling image is much clearer and hidden details came to life, resulting in an image that is a little closer to 3D rendering . In this video, Stable Diffusion, which has recently become a hot topic, is tested for game icon production. Thank you for watching the video.
