Hi there, this is one of the beginner use cases and I learned it from one simple Youtube Video. The only prerequisite is to be patient with the google colab free subscription or buy a subscription so that you can access the ‘L4’-GPU in coolab. Otherwise no python knowledge required, just press two buttons to reach the AI image generator graphical user interface (GUI). I do really like the interface, it is already sophisticated, but those buttons that are not so relevant are well hidden inside ‘advanced’ pull-down menues. For this experiment, I tried to keep the original as close to the generated anime cityscape, and I have I succeeded just partially. The main problem is that the sun shines down vertically and the shadow lines from the roofs on the side turned out to be often interpreted as a narrow field where grass and plants are drawn, or alternatively just draws a pavement. It also draw the shadow right, and this changed for each parameter setting, which I say later more, but then there emerged another problem that always tiped the balance on the other side: The house, particularly on the left, was generated with reddish-orange walls. See for yourself:
This is roughly what stable diffusion generates when you simply chose the SAI anime style, without changing the whole model to the ‘anime’ setting. This can be done through
So you can see, everything is kind of in orange-red, but the shadows on the alley way are neatly aligned our goal.
Here we got the closes’t representation of what is in the image. The end of the allay sticks into the eye and the shadow lines are perfectly mirrored from the original picture, however, you already have the feeling there is something wrong, and its about the impressions, the feeling that the overall picture give the viewer is completely off. While the original picutre looks like a warm summerday, the right looks like a winter day. So also this picture isn’t mirroring the real world picture.
Another great generation, but the orange really doesn’t fit the original, so we can make a nice mixed-image, vizualising the cornerstones that generative AI has reached.
Okay, AI is pretty good, but it seemingly needs more human guidance to understand its task, since you and I clearly judge the generated pictures as esthetically profound but sparking very different feelings.
With the inpainting, we can fix the orange area, but the pictures wholeness decreases. If you compare the leaves ot the left tree, you can see that the inpainted area looks somehow different.
Here I show you some other pictures, which just show the diversity has been learned by the anime-style image generatior tool.
