“An A.I.-Generated Picture Won an Art Prize. Artists Aren’t Happy,” announces a recent New York Times article. The reporter goes on to say that, at a state-fair art competition, one artist won the digital-art category for an AI-generated image.
Although I have no qualms about the artist submitting this piece, especially since he was up-front about the AI generation, I found the idea that computers might replace artists disturbing. That and curiosity led me to try some of the available art-generation apps.
Frankly, I gave up on all but one due to complicated instructions or poor results on initial demos. However, I did spend several hours on Midjourney, an arm of Discord, and attempted to create an illustration for an upcoming blog post.
As I worked, I realized that Midjourney is basically a search engine that attempts to locate an image already in existence. It then morphs the image somewhat, using a painterly style that blends elements to suggest form, but that produces a picture in which no form is clearly discernible.
When using Midjourney, the more one’s description matches an image already in existence, the better the results. For example, one creator requested an image of cavemen looking at a cell phone. The result was charming, but the app simply appeared to locate a picture of men looking at a cell phone (how many of those have we seen?) and then added Neanderthal features to their faces.
Another requested gloomy pictures of black crows. The results fit the request, but there are many gloomy crow pictures already online.
A major weakness with Midjourney is that it does not recognize verbs, so it is impossible to create an illustration of characters interacting. While experimenting with the app, I used common verbs such as chasing, running, and pushing, all of which were ignored. And when two or more characters were mentioned, the app attempted to morph them into a single creature.
Below are examples of descriptions and results:
My description, or prompt: A tired, sweaty hiker stands on a Florida hiking trail. She pushes her walking stick in deep, black mud. There are tall grasses behind her. In the grass lurks an alligator and a pit bull. Both animals look at the hiker. In the foreground is a broken compass.
This is the best result of four options provided. It caught the words [hiker], [Florida hiking trail], and [tall grasses] and located a matching image already online. It ignored all other information in the description.
But really, I was hoping for something whimsical, and the alligator was a critical part of the illustration. So I tried with this prompt: Colorful cartoon. Hot, sweaty, female hiker. Knee-high in black mud. Surrounded by high grass. An alligator and snarling pit bull in the grass.
Again, the images generated don’t fit the description. None of the hikers stand in mud. Only one image shows high grass. Is the black blob an alligator? Maybe it ate the pit bull?
Now I’m feeling annoyed and try something different.
Prompt: Female hiker wrestling an alligator and a pit bull.
In these images, the app appears to morph the dog and alligator into a single animal. And the hiker in the upper right-hand image appears to have a tail that’s part alligator and part pit bull.
Prompt: Funny cartoon. Female hiker with a leashed alligator. Colorful.
Midjourney attempts to merge the alligator and hiker into a single being.
Prompt: Female hiker walks over a trail full of alligators.
Umm, where are the gators? Oh wait, a black blob follows one hiker, and in another image, an alligator has disguised itself as a hiker.
Prompt: A female hiker steps across a trail covered with alligators.
Finally something somewhat usable. However, look at those gators. None of them have heads. There is only the suggestion of gators.
Prompt: Cheerful female hiker on summer day walking a trail filled with alligators.
Does the app even know what an alligator looks like? I decide to test it.
Prompt: Detailed alligator on muddy hiking trail. Florida.
Afterwards, I tried two completely different subjects (not shown here) with equally poor results.
Artists, I can assure you that you are still needed. AI is a long way from being truly useful, and it lacks thoughtfulness, appropriateness, and creativity.
Although computers can provide a definition of a verb, AI doesn’t understand the meaning offered, which means it can’t put things into context. It can’t put things together in a meaningful way, or imagine a story from a given description.
AI is simply a tool and, at this point, a poor one. And since AI image-generation apps appear to simply copy and morph images already in existence, it’s easier to search for pictures on sites like Pixabay, which offers copyright free material. One can then credit the original artist or photographer because, for us humans, those little acknowledgements make a difference. They’re recognition that we’re all interconnected and depend upon one another; that humanity still matters. Each of us is as unique as the day’s sunrise.