Adventures with automated art

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

Do you desire to make art, but can't draw? I'm like that. My father painted as a hobby for much of his life, and my house and my brothers' houses contain numbers of his paintings; he also sold a few and gave away even more. One of my brothers got all the art talent, and I got very little. I learned of prompted art generation almost a year ago, and last November I decided to give it a try.

I tried out all three major art-generation sites, DALL-E2, MidJourney, and Stable Diffusion. My early take on these is posted here. Since that time I've used DALL-E2. A quick intro:

• When you start you get 50 free credits. Another 15 free credits arrive each month thereafter.
• Added credits can be purchased at the rate of 115 for $15. I find that I buy more credits about every second or third month.
• To start, after setting up an account and logging in, there is a space to enter a prompt. You'll use many of your first 50 credits getting used to writing prompts. There are tons of YouTube videos where you can get creative and useful ideas.
• You can also upload an image you want to edit. I won't get into that.
• When you click Generate, you get a panel of four square images. Their size is 1024x1024 (in the earlier version of DALL-E they were 512x512).
• You can click on one of the four squares to see it larger, and download it as a PNG file; its size will be 1.5-2 Mbytes.
• When looking at one of the squares you can click Edit. The two main options here are Erase and Extend (called Add a Generation Frame). You can also upload an image at this point, but I am not sure what happens; I haven't tried it.
• Use Erase to remove a part of the image you'd like re-rendered differently. Then click Generate and you get 4 versions to look among (a 4-dot clicker to select). You can Accept one of them or Reject and try again.
• Use Extend to activate a 1024x1024 marquee that you can move around to an overlapping area. Be sure there is some overlap, or you'll get a discordant appendix. Here again you have four versions to select among, to either Reject or to Accept the one you like best.
• I do tons of Extending, to turn 1024x1024 images into 1920x1080 or larger. I haven't tried going all the way to 4k (3840x2160); that would take at least a dozen Extend tries.
• Note that each time you click Generate it uses one credit, whether you Accept or Reject the result.

The two images below are the beginnings of a Mountains theme. The prompts that generated the starting squares were

1. Mountainous Landscape in the style of the Hudson School
2. A mountain cabin in the style of Thomas Kinkade

These are the first and fifth "paintings" I generated. (You can click on any image for a larger version; this image is a 2-pic pasteup). Each took six or seven Generation events, so each cost less than $1. For the second image I found I had to edit the prompt after a couple of steps, or DALL-E2 would insert more cabins and even whole floating mountains. Prompt editing is a skill in itself. It is part of the "negotiating" one does with DALL-E2 to get the results you want. So is Erasing sections, small or large, and using up more Generate credits.

For my second painting I did a Still Life in the style of Paul Cezanne. Later I did three others. All are shown here, and their prompts, clockwise from top left, were

1. Still Life with apples, pomegranates, and bananas in the style of Paul Cezanne
2. A still life painting with cheeses and fruits and a vase in the style of Clara Peeters
3. A still life painting with peaches, grapes and a vase of flowers, in the style of Raphaelle Peale
4. A still life painting with a cornish hen, grapes, and leafy greens in the style of Pieter Claesz

For a sense of the scale of extending these images, here is the initial panel from the Paul Cezanne prompt:

This is shown at a roughly similar scale. You can see that I Extended the third square. For all the final paintings, the size was somewhat greater than 1920x1080, and seldom had the 16:9 ratio. Thus I could crop most of a painting to the 16:9 ratio, leaving out the color block DALL-E2 puts at the lower right of all its images.

As I got more experience with DALL-E2, I learned that in addition to asking for styles of a certain artist, one may also specify a genre, such as "painting", "digital art", "pixel art" (blocky), "3D render", "clip art" and others. I'll show examples of some of these in future posts.