Retrospective: 2 years of SI art

In late 2022 I got access to Dall-E2. After a little experimentation, I entered the prompt, "Still Life with apples, pomegranates, and bananas in the style of Paul Cezanne". Of the four images offered me, I put the one I liked best into the Edit mode and "outpainted" the square image to produce a larger, rectangular one about the right size. It looks a lot like a painting Cezanne might have done; if you go to Google Images and enter "Cezanne still life" you'll see a great many examples. This image was clipped to 1920x1080 from the 1984x1280 original; I signed it on behalf of myself and my co-creator. It is dated 2023 because I didn't crop or sign it until January.

Today I decided to see how successor programs are doing. To keep this from getting overly long, I'll drop in a paste-up of 12 images from three of my favorite SI (Simulated Intelligence) art agents, and discuss them briefly.

You can click on the image to see it larger, which makes the text (file names) easier to read. This is a screen shot of Windows Explorer with "View/Extra large icons" set.

• The first two are from Dall-E3, the successor to Dall-E2. I use it through Bing, where I can select a square image (1024x1024) I like and have it re-rendered in wide mode (1792x1024) which is very close to the 16:9 ratio I prefer. More on this in a moment. Although these are a lot like Cezanne works, they are more cluttered than most of his still life paintings. In closeup, they are more detailed and "clean" than the Dall-E2 version, as are most of the rest.
• The next three are from Playground AI, using the three "engines" currently available, Stable Diffusion XL, Playground v2.5, and Playground v3.0 (currently Beta). Ordinarily I get the "brightest" results with PG30, but this time all four images it offered me were washed-out in this way. Perhaps code has been added to avoid making images that are too much like works by famous painters.
• The other seven are from Leonardo AI, which has ten engines, and each has 10-20 or more modes such as "Dynamic", "Vibrant", "Illustration", some of them unique to a particular engine. The engines I used here are Leonardo Lightning, Portrait Perfect, Illustrative Albedo, Concept Art, and Lifelike Vision. The last image of this set looks the most like Cezanne's work, in my estimation; it was made with the settings "Lifelike Vision/Retro". The ones made with Illustrative Albedo and Concept Art are too bright, while the others have a pretty good "Cezanne look."

I prefer the term Simulated Intelligence to Artificial Intelligence. I think of these art generators as tools that enable me, someone who gave up drawing in frustration at age ten, to think of a picture I would like to see and "negotiate" with a machine intelligence, which has no ideas but can draw or paint with great skill. It is akin to an author who collaborates with a coauthor, who doesn't have the original ideas but has writing skills that supplement or complement those of the lead author. For the record, I don't try to use ChatGPT or CoPilot or other tools to "help" me write because I think I write well enough—I've surely worked hard enough on it!—and I like to keep control of my "voice".

I also produced an image using Gemini (formerly Bard), which has a new image generator (I haven't tried it on humans). I produced one image, and even though I asked it to render a wide format image, I got a 2048x2048 square, but it has plenty of head- and foot-room, so I can crop it to wide format.

I have two main uses for art generating agents. Firstly, when I write a blog post (usually a book review) in https://polymath07.blogspot.com/, I like to include an image, sometimes from the book or from the Author's web site, and other times from an impression I had while reading. Generated art is not copyrighted, which saves me gathering credit info. Secondly, I devise prompts for images I would like for my private image collection. These are intended to run in a screen saver and I use some for Zoom backgrounds. I have 1920x1080 screens, so I make them that size or larger, as long as they have a 16x9 aspect ratio.

The images produced by these programs are often close but not quite exactly 16x9, so I have to adjust them. They are usually smaller than 1920x1080, so in most cases I use Upscayl (the desktop version) to double the size. For example, the "16x9" setting for most of the Leonardo AI engines produces images that are 1368x768. 1368 is not divisible by 16 but by 8. However, when I use Upscayl I get 2736x1536, and 2736 is divisible by 16. But then 2736x9/16 = 1539. I have to crop the other way, so I need to cut back to 2720x1530. I use IrfanView to do this. I sign the cropped image and save it in my Screen Saver folder.

For an image that is almost right already I don't have to do a full doubling, and I can use IrfanView to increase the image size a little. For example, the Dall-E3 images are 1792x1024. I start by cropping to 1792x1008 (112 times 16x9). IrfanView produces a good result upsizing that to 1920x1080, a factor of 15/14.

It is satisfying to produce artwork to illustrate my ideas.