Hello In this video, I will experiment with making game skill icons using Stable Diffusion . In the previous video, I made icons for items with specific shapes, such as apples, books, and shields. Based on a stable composition, it was important to reenact with the same completeness. Skill icons, however, are a bit different. In addition to basic representation, elements such as dynamic composition, symbolic color, and intense expression that can stimulate the user's imagination while revealing the characteristics of the skill are required. So, let's take a look at how we can create these skill icons with Stable Diffusion. As before, we use the Stable Diffusion WebUI Image2Image function. Looking at the other settings, I will first use revAnimated_v.121 for Checkpoint , Sampling Method: DPM ++2M Karras Sampling Step: 30 Batch Count: 4 CFG Scale: 15 Denoising Strength: 0.5 I will fix those values and work. I copied the Prompt and used it as it is. This time, I referred to a little bit of documents published on the web, so I added only those that were deemed necessary . 3D Rendering, Signature, Text I'll put it like this. The first thing I'll make is the icon in which a hand covers bright light. Based on the results of the last experiment, I prepared an image that contains basic composition, shape, and color information. As expected, the color, shape, composition, etc. were well maintained, but the rim lighting effect caused by the bright light was barely noticeable. The image was too flat and did not feel as dramatic as I expected. Rim lighting was added to the outline of the hand using a simple layer effect in Photoshop. Let's create an image in this state. The rim lighting effect was clearly added around the outline of the hand. However, the superficial appearance of the hand did not change significantly. The black area in the image didn't seem to convey any information during the image creation process, so this time I decided to add color information to the completely black area. I put a very dark blue on the fingertip and gradually It made the skin color appear. The detail and volume of the wrist part where the color of the skin was added was much improved. However, the finger part applied to the dark blue color showed an unexpected look. I was able to get a satisfactory result by modifying the color of the starting image a few times. As the level of description of the hand rose to a certain extent, the 'Bright Light' felt relatively disappointing. I decided to add some visual effects to emphasize the feeling. First, I added information for the visual effects to the starting image, and then I re-adjusted the color of the hands so that these visual effects could be reflected in the hands as well. And I added Bokeh and Lens Flare in the prompt. Satisfactory results were obtained at the level of color, shape, composition, and description. However, at the same time , unwanted elements such as rings, bracelets, and sleeves were added to the image . I added these elements to the Negative Prompt and created an image once again. I was able to achieve the desired result. The second set of icons to be produced is a figure of a person standing in front of strong flames. This time, I prepared a starting image with the Rim Lighting effect added from the beginning. For reference, the shape of the person shown here is an image created online. In the prompt, I simply added the word Flame, Man. Let's create an image. The color and shape composition came out as expected, but it was a little disappointing that the flame behind the figure was too weak . I decided to provide noticeable clue to it. I created an image. The approximate shape of the flames in the starting image is depicted in a specific form. However, there is one peculiarity here. Despite not touching other parts , the detailed shapes of the person's face and clothes It's starting to appear. It seems that the shape of the newly drawn flame has an effect. I'll correct this part in Prompt. And it's good that the figure is standing in front of the flame, so it's dark, but because it's too dark, the image looks flat as a whole. The area of the figure I'm going to add a bit of color inside so it acts like fill lighting. I'm going to modify the starting image so that the chest area is lit with a weak blue light, and I'm going to add faceless and topless to the prompt. The face has disappeared and the upper body is not wearing clothes, and the blue color that was painted on the chest of the person in the starting image serves as fill lighting, so you can see that the overall sense of volume is much more alive with the level of description. The third icon is a hand touching the bloody surface. Based on previous experience, I prepared a starting image. I expressed the skin color with weak rim lighting. I will simply add information such as Hand and Blood to the prompt and create an image. The color and texture of the illuminated hand by Rim Lighting were expressed very well. However, there is one unfortunate part here. It is the shape of the hand. There is a strange shape of the part visible between the fingers . It seems that there is a limit , so I decided to try a function called ControlNet. ControlNet is used when generating images in Stable Diffusion. It is an extension function that can add various conditions. For example, in the case of Canny, the outline is extracted from the image. Then, based on the outline, only the inside or outside of a specific object can be used as an image creation area. So, I will use the method using depth map. There was an image that I used to draw the shape of the hand in the first starting image. Fortunately, that image was in the form of a depth map, so I will use that image as it is. I opened that menu and inserted an image with depth information. PreProcessor is used when extracting an image with depth information from a normal image . Go to the existing model and select a model that uses depth information. Lastly , check Enable to use ControlNet. Let's create an image . If you compare it, you can see the difference. You can see that the additional depth information provided through ControlNet solved a lot of the strange shapes between the fingers . The hand shape problem was solved to some extent, but one problem arose: the spatial relationship between the hand and the blood in the background . It's an ambiguous spatial relationship that's hard to even tell it's far away. Looking at the image I used for ControlNet, the hand has enough depth information, but the rest of the area is completely black. This seems to be the cause of the problem. In order to create a depth map that clearly contains the relationship, modeling data of the hand is required. However, I don't have the modeling data. I created a new image with depth information using Photoshop. After applying this image to ControlNet, the image was created. Unlike the images created just before, you can see that the fingers are firmly pressed against the floor. The last icon to be created is a setting in which light green smoke or aura surrounds the skull. This time, I created an image with depth information myself using Blender and a free skull modeling data. I applied this image to ControlNet. But, as you have seen before, if the brightness changes abruptly in the image with depth information, the background and foreground objects are clearly separated by that much . I think I need to tweak this map a bit because it's a setting that the skull is in the middle of the green fume. With the ControlNet's PreProcessor I'll create an image with depth information. And in Model, I'll select a model that uses depth information. Now I'm loading the prepared starting image. It contains all the information for lighting effects, etc. Lastly, let's add Skull, Fume to the prompt. Let's create an image. An image that reflects the concept and content well is created. You can see that it is merged well in the fume. The image itself came out well, but looking at the finished result, there is something a bit disappointing. Looking to the left of the icon, there is a rim lighting representing the outline of the skull and bright yellow-green smoke. To add more contrast I will prepare a starting image with the lighting direction changed and create the image again. Let's change the starting image to create an image that emphasizes the outline of the skull in the dark as if changing the lighting direction. In this video, we looked at the creation of game skill icons using Stable Diffusion 's Image2Image and ControlNet. Thank you for watching. .
