DALL-E is an online tool that uses machine learning to generate digital images from plain English text descriptions. You type a description of something real or imaginary and the program does its best to create a unique image based on that description. After some time on a waitlist, I recently received an invite that allows me to create and download a limited number of artificial intelligence generated images per month. This came at a good time, as I recently found a watercolor print of the Duluth hillside in a Lincoln Park shop that I liked quite a bit but could not afford. I decided to use some of my AI image credits to see if I could get the automated system to produce Duluth art of at least somewhat comparable quality. In the examples that follow, I describe this process, showing what worked and what did not. The captions of each picture show the text query that generated the image.
Trying to get the program to produce specific architecture or just asking for the image to be set in Duluth did not work very well, as the program blends elements of its source images. This means that trying to add a specific Duluth landmark results in an image that seems a bit off. Or In the case of the image below of a family in front of the Aerial Lift Bridge, very off and a little terrifying.
After a few unsuccessful attempts of trying to get Duluth images by including the word Duluth or well-known Duluth places in the query, I started trying to break Duluth down into its most basic elements: a city, a hill, a lake, trees. This started working a bit better.
Note: For each query, the program generates four outputs. For the examples here, I have chosen the output that best matched what I wanted.
The program allows you to imitate a specific painting medium, such as oil or acrylic, but you can also create an image in the style of a particular artist, like Edward Hopper.
I tried adding in more specific elements to see if that would give the outputs a better sense of place.
Changing just a few elements of the description often resulted in a completely different output.
At some point, I realized that specifying the color of the ship improved the Duluthiness of the results dramatically.
The challenge was keeping the query manageable while to trying to invoke specific places, like Park Point.
In addition to artwork, one could potentially use the program to create fake historical photos like a trained black bear skiing at Chester Bowl …
… or to generate images that mimic popular Perfect Duluth Day posts.
But these images, while amusing, don’t match the prompts exactly and aren’t particularly photorealistic. In addition to functional issues like not knowing what a loon is, the program tries to prevent potential misuse by scrambling any human faces in the image.
A better use might be to quickly generate images that show possible futures. For example, what downtown Duluth might look like if Essentia keeps expanding.
Of course, paintings and photographs aren’t the only possible outputs. The program will also attempt to generate an image for any object you describe.
Most of my queries, however, remain directed toward trying to create some Duluth art for my wall, like this attempt at a watercolor painting of a bridge over the Lester River.
None of these AI-generated images are quite as nice as the print that I found in the Lincoln Park shop, but I remain hopeful that something incredible is just one well-formulated query away.
If you think you know what that query might be, feel free to put it in the comments and if I still have some credits left, I’ll give it a try and post the result.
Leave a Comment
Only registered members can post a comment , Login / Register Here