It has been a dizzying three months or so since artificial chatbot service ChatGPT was unleashed on an unsuspecting public. However, it has been much longer since the same ownership company, OpenAI, released its artificial intelligence artwork applications under the DALL-E (think Salvador Dali) branding.
Wherever I’ve been over the past few months I have made a point of demonstrating these services, usually eliciting looks of disbelief from those who see these applications at work. I expect to be similarly demonstrating others, such as Google’s Bard and Microsoft’s version of ChatGPT integrated into its long-languishing browser, Bing.
At a recent morning coffee gathering, I happened to sit next to a noted B.C. artist, calligrapher, and iconographer, John Suter, profiled in this paper back in 2018. I will be the first to admit I know very little about art and certainly could not paint anything better than a trained elephant might.
So when I began my demonstration for Suter I initially only focused on the text service, ChatGPT, asking for innocuous things such as providing a brief history of impressionist painting, and writing me a story about iconography. My artist friend was impressed, amazed even, so I decided to press on and demonstrate DALL-E.
I wasn’t really sure how to proceed and decided to show him the rudimentary examples I had produced with DALL-E. I pointed out that, unlike ChatGPT, DALL-E restricted users to 15 free creations a month and that fortunately my count had reset that morning. My previous experience with DALL-E is that I typically exhaust all 15 credits in a single session, so addicting can this service be.
DALL-E, when launched, offers two suggestions for a first creation, either “Surprise me,” or “An Impressionist oil painting of sunflowers in a purple vase.” Not wanting to appear too ignorant of the art world, I showed Suter my first creation with DALL-E, produced with the words “Van Gogh style painting of a large ceramic vase with red roses standing on a teak table.”
Let me reiterate, Suter is a very experienced artist. He was dumbfounded to see the four pieces of art created by DALL-E from those instructions. That led to a discussion about ownership of such artwork, copyright if you will. We will leave the result of that discussion to another day, mainly because there was no clear outcome. The concept of copyright in a product created by an artificial intelligence application is not clear at all. After all, the AI application was trained on existing artwork.
After showing my friend some additional examples that I had “created” through DALL-E I decided it was time to put him to work. He could see from my simple examples that more detailed textual descriptions produced results more akin to what a user might have in mind.
Here’s what Suter decided to try for his first effort: “landscape lake in front, mountain in back, blue skies, rowboat on lake, ducks in background, houses on mountain, and a castle.” I wasn’t sure how that would play out, but within 15 seconds up popped four versions, each meeting, mostly, the criteria spelled out.
“Wow!” my friend exclaimed. However, we immediately spotted that a key word had been omitted from the instructions, namely “painting.” These outcomes were all photograph-like. Whether they were a single photograph that DALL-E had simply “borrowed” or whether these were conglomerates from multiple sources was difficult to immediately discern. A Google search on the same instruction set produced nothing remotely similar to what DALL-E had “created.”
Before adding the word “painting,” we tried the “regenerate versions” option, which gave us a total of eight variations on the original text specification.
Running DALL-E this time with “landscape painting, lake in front, mountain in back, blue skies, rowboat on lake, ducks in background, houses on mountain, and a castle” caused the application to buckle. It stalled around the 90 per cent completion mark. We tried again. Same problem. But on a third try a result set appeared.
Accompanying this column you will find examples of DALL-E’s output. DALL-E has essentially followed the instructions almost to a tee. One variation to be seen is that the ducks, or other birds, are mostly in the foreground rather than in the requested background.
Want to try DALL-E? Go to openai.com/dall-e-2 and sign up with a valid email account. You can have DALL-E create images from scratch or make edits to existing images, all with natural language instructions. As for copyright on these images, let’s just say again, that’s a little murky for now.
Follow me on Facebook (facebook.com/PeterVogelCA), or on Twitter (@PeterVogel)