Automated western art, ABCD test

 kw: experiments, artificial intelligence, simulated intelligence, art, generated art, images, comparisons

Several months ago I had the idea to try more prompts that differed only in the style requested. I ran four sessions with the prompt

Painting of a western landscape with a dramatic sky in the style of [Artist]

The four artists I chose were Albert Bierstadt, Thomas Moran (I was familiar with both), John F Kensett, and R Brownell McGrew (not strictly western, but prolific landscape artists). The initial squares I chose from what Dall-E2 generated for each artist are these:

I outpainted them to these sizes: 3200x1984 (first 2), 1880x1984, and 2944x1920. Resized for composing together, and reduced for this montage, here they are:

The artists are as stated above, beginning at the upper left and proceeding clockwise. With a little study you can see where the original square would be in each image. I cropped large, not quite maximal, portions of each image to a ratio of 16:9, with these results:

For three of these I emphasized the sky, and for the one in Kensett style I emphasized the desert floor, while still leaving plenty of sky.

I hunted through public images by each of these artists so I could show to what extent Dall-E2 emulated him. I searched for paintings that might bear some slight resemblance to the generated ones. This montage is of quite reduced versions of a painting by each artist, clockwise from top right:

1. Oregon Trail by Bierstadt
2. Green River, Wyoming by Moran
3. Mount Chocorua by Kensett
4. Winter Morning by McGrew

Mt. Chocorua is in Hew Hampshire; JF Kensett is considered a Hudson Valley School painter.

It seems that Dall-E2 can produce distinct styles, but not quite distinctive styles-by-a-particular-artist. Given the millions of high quality images of both photos of wonderful scenery and classic art available online, I don't really need to generate good looking landscapes. This was a fun project, but not something I intend to to a great amount. True (human) artists have little to fear from generated art.

In the future we can expect tools like Dall-E2 to get better and better, and I expect within a year or two the ability to generate highly detailed, 4K images in a single Generate (maybe for a higher price per credit), which can then be edited, including with pasted-in bits one could upload during the process. Will they have a "human touch"?

Why would we want paintings by a nonhuman entity to have a "human touch"? It's the nonhuman touch that interests me. The "Alien City" images I produced earlier are intended to be "not human". Even when I asked for a picture "in the style of Chesley Bonestell" I didn't expect slavish imitation, and I was right about that! The image that resulted is the "least human" of the four.