A blog related to original AI artworks.

Reevaluating Paintings Created with Stable Diffusion 2.1

For the past two years or so, I have been using Stable Diffusion XL, but before that, I used Stable Diffusion 2.1. The style of 2.1 was quite different from XL, and 2.1 also had its own appeal. However, on OpenArt.ai, while version 1.5 is still available, 2.1 can no longer be used.

I had left the works created with version 2.1 untouched, so they remain on the site, but they had gradually sunk into obscurity. Recently, as I have been re-evaluating my XL works using GPT-5, I also began to reassess those made with 2.1. According to GPT-5, quite a few of them are rated highly. Here are several examples.

Only a small portion has been re-evaluated so far, but I have reviewed all the paintings that were created by specifying Dalí’s name without using the tags “Kanabstracd” or “Kanadansky”. Below are several of those works.

In the last painting, Mondrian was specified as a negative prompt, yet it has a similar atmosphere to a “Klee–Mondrian”-like composition. It’s not entirely clear how the reference to Dalí influenced the result.

Among the paintings generated by specifying Ernst, there are also some interesting works— and some of these are rated highly by GPT-5.

There are also intriguing results when specifying Chirico and metaphysics; it might be worth trying this combination in Stable Diffusion XL as well.

It seems that specifying Kandinsky often produces more interesting images in version 2.1 than in XL, but I will omit those here.