zsx

zsx

先作为主站备份 源站：https://my.toho.red
HomeArchives

AI Painting Experience - Generating Corresponding Virtual Image Using School Emblem

#随笔/记事47
AI Translation
This post is translated from Chinese into English through AI.View Original
AI-generated summary
The article discusses the popularity of AI painting and the author's experience using their school emblem to generate a virtual image. They mention using specific AI models and techniques to improve the quality of the generated images. The author shares their process of refining the image and seeking feedback from others. They ultimately choose a final image, although they acknowledge some flaws, particularly in the eyes. The article concludes by reflecting on the development of AI in the field of painting and its potential impact.
Since the release of the AI drawing feature by NovelAI, AI drawing has suddenly become the focus of attention across the internet. The reason, of course, is that the AI drawing style has finally moved away from the abstract era of the 6pen and become a tool that can consistently produce ~~waifu~~ works. With such a great ~~waifu~~ drawing software, it naturally piqued my interest, so I followed the trend and used the school emblem to generate corresponding virtual images.

Disclaimer: The virtual images in this article are all created using the stable-diffusion-webui model released by automatic111 and the waifu-diffusion model released by hakurei. The results may differ from using NovelAI.

First, let me share a spell (because for some reason, I won't be sharing the school emblem for now):

{best quality}, {{masterpiece}}, {highres}, original, extremely detailed 8K wallpaper, 1girl, {an extremely delicate and beautiful},solo,straight-on,upper body,depth of field,east asian architecture,grey_hair/grey_eyes,looking at viewer,standing,school uniform,oval face, Negative prompt: lowres,bad anatomy,bad hands, text, error, missing fingers,extra digit, fewer digits, cropped, worstquality, low quality, normal quality,jpegartifacts,signature, watermark, username,blurry,bad feet, Steps: 20, Sampler: Euler a, CFG scale: 7, Seed: 831186504, Size: 1024x1024, Model hash: e1de58a9, Denoising strength: 0.75, Mask blur: 4

Here is the result of the first attempt at alchemy:

Enter image description here

At first glance, it looks pretty good, right? That's what I thought at first too, but when I presented it publicly and received guidance from a classmate who seemed to be a painter, I discovered the following shortcomings:

  1. The face is a bit wide and the proportions are a bit off. This problem can be easily solved using the Liquify filter in Photoshop.
  2. The spacing between the eyes is also a bit wide, but the same solution applies.
  3. There doesn't seem to be any immediate issues, but it just looks strange no matter how you look at it. This is probably a common problem with AI drawing, and since I'm not a professional artist, I'm powerless to make any modifications.

So after some adjustments, the first and second issues were successfully resolved, resulting in a more acceptable appearance:

Enter image description here

Even after the modifications, it still feels a bit imperfect, so I decided to put this image into img2img to see what it could be iterated into. Unfortunately, it seems that luck wasn't on my side, as each summon resulted in something indescribable. Just as I was about to give up on iteration and use this image as the final product, I came across a special image:

Enter image description here

Fortune favors the prepared mind, and after generating over 100 images, I finally iterated into a waifu relatively high-quality one. It turns out that Chinese waifus have something in common. After the public presentation, this image successfully surpassed the previous one and became the next candidate.

At this moment, I suddenly discovered a technique for generating images: if I use a resolution of 1024x768, the actions and backgrounds are generally more rich compared to a 1:1 resolution, seemingly because a longer image can accommodate more information. So in the subsequent iterations, I intentionally used a resolution of 1024x768 and obtained the following 4 waifu relatively high-quality images:

Enter image description here
Enter image description here
Enter image description here
Enter image description here

The competition for these images was more intense after the public presentation, and the popularity reached its highest record of the day. Although the final scores were similar, after my secret manipulations comprehensive judgment, I selected the second image as the final product.

The overall appearance of the second image does indeed match my waifu expectations, but I had some difficulty with the detail of the mouth: although it would be cuter to have an open mouth, I felt that it would be more elegant to change it to a smile. After weighing the pros and cons, I ultimately chose to change it to a smile:

Enter image description here

This seems to be the final product, right? I still feel that there are some obvious issues, with the biggest problem being the eyes: AI doesn't seem to pay much attention to details, and if you look at the eyes from a distance or for a long time, they appear lifeless and may even induce the uncanny valley effect. The original plan was to have a professional artist make further modifications after the AI generated the final image, but it seems that the art community has a strong bias against AI, as I couldn't find any willing artists to make the modifications despite going to several places, and I even received some insults. So this is how it will be, this is the final product, and since it's just for fun and not for any subsequent operations, I don't need to consider it further.

The development of AI has indeed exceeded the expectations of most people. It was previously said that jobs requiring creativity would never be replaced, but today, one of those jobs, drawing, is already facing the threat of being replaced. However, let's save the deeper discussions for later, as we are just ordinary users. Setting aside the controversy, the results of AI drawing are really amazing, aren't they? They can satisfy any kind of waifu, truly remarkable.

2022.10.11

Loading...
Ownership of this post data is guaranteed by blockchain and smart contracts to the creator alone.