日本語ページ   ▶AI art top   ▶AI art blog   ▶Image top  

🤖🖼 AI Art Blog 🅰🅸 🅰🆁🆃

A blog related to original AI artworks. I make AI artworks using coined words and emoji as prompts.


Diversifying Generated Images by Specifying Painter Names in Positive and Negative Prompts

2023-05-13 15:00

When the same painter name is specified for positive and negative prompts, Stable Diffusion 2.1 often generates very interesting images. And sometimes, very diverse images are generated. Specifying the name of an artist in a negative prompt to obtain diverse images may be related to the fact that an individual artist creates diverse works by negating their past creations.

ta8obwfh 0p396zlu

I tried specifying the same word in positive and negative prompts for the first time in Stable Diffusion 1.5, but it didn’t result in anything particularly interesting. I tried using painter names such as Picasso, Dali, or Bosch. The following image was obtained by negating Picasso, that is, providing “Picasso” as both a positive and negative prompt.

r6aitp8h jb3nokui

It is not bad, but I don’t think it is a particularly interesting image. I tried others as well, but I didn’t find any images more interesting than these.

Also, the following images were obtained by negating Bosch.

zcgtdu2w 1mlcuwyk

They are a bit chaotic, and it is hard to tell what they are depicting.

Picasso and Bosch both have some interesting aspects, but they are not particularly captivating. Dali was a bit more interesting.

_8_z0q2r f4pi0ykk

However, even these were not particularly interesting.

What changed the situation was trying it in Stable Diffusion 2.1. I think the method of balancing positive and negative prompts worked much better in SD2.1, but the fact that the number of generated images was small when I tried it in SD1.5 also had an effect.

So far, no particularly good images have been drawn with Picasso, but when I tried SD2.1 with other painter names, I got interesting results. Kandinsky was particularly interesting, but I will write about that in another article. Here, I will first talk about Bosch.

=2yfe4q9 g5p1f64c

I think it has more unity overall (that is, it is not as chaotic) than Bosch-style paintings generated by SD1.5. Compared to when no negative prompt is specified (whether it is good or not), the colors are richer.

I haven’t tried it with Dali yet, but I have with Seurat, Chagall, and Van Gogh, among others. Let’s start with Seurat.

0bw1gekd cwmglb4m

Seurat’s painting of people at the beach is famous, but the painting on the left takes up the same theme while the coastal landscape and color scheme are completely different. The diversity is appealing, but the people are not drawn well and the sky and sea look rough. The painting on the right, of a park, is similar. The colors are attractive but it looks rough. There are other similar paintings, but they are omitted here.

The following images are when Chagall is specified.

9lthfldh vtmucxbi

There’s nothing particularly interesting, but I think the peripheral and related elements in Chagall’s paintings are brought out.

I also tried with Van Gogh, Severini, and others. Because they worked better than with artists like Seurat or Chagall, I will write about them in separate articles. In any case, as mentioned earlier, specifying the name of an artist in a negative prompt to obtain diverse images may be related to the fact that an individual artist creates diverse works by negating their past creations.


Dasyn.com デイシン
Created: 2023-05-13 15:00   Edited: 2024-03-28