Anthropomorphized Space Needle with a conical helmet covering top half of face.
An arrow pointing leftHome

I tried to make AI art generators anthropomorphize PNW landmarks

  • Justin Caffier

“As improved as this era of AI art generators is over the last, it’s only an incremental one.”

It’s easy to understand why the flesh-and-bone visual artists, like those who lost to Jason Allen’s AI-generated work “Théâtre D’opéra Spatial” at the recent Colorado State Fair, are salty. As we graduate from the stone to bronze age of AI-generated artwork, these artisans are seeing their lifetimes of passion, dedication, and practice out-performed in a few seconds by unfeeling lines of code, so it’s easy to see why they might be sweating and feel their job security is under siege.

Though I feel for these artists, I don’t share their same degree of concern. I’m not a visual artist by trade, so I can see the inherent value in creating a new avenue for allowing people who might have otherwise not have the time, technical skill education, or physical capabilities needed to produce such works the traditional analog ways. Just as GIMP and Photopshop opened up the door into the art world for a generation before, AI-generated art may give a creative voice to an era of machine-learning Michaelangelos.

But the big reason why AI’s threat to trad artists seems overblown is just how incomparable to the arresting power of the human-made pieces any of the text-prompted stuff is. Sure, a few state fair judges deemed Jason Allen’s piece worthy of the blue ribbon, but have you seen the human-made stuff that state fair judges have chosen in years past? We’re not exactly dealing with a gaggle of Jerry Saltzes here. To the rest of us, “Théâtre D’opéra Spatial” came off as a forgettable circa 2009 HD desktop wallpaper. And that’s not a ding unique to this case.

As improved as this era of AI art generators is over the last, it’s only an incremental one. There is still a noticeable and diminishing patina of “digital art unseriousness” over all the produced work we’re seeing. And no qualifiers about having the programs create the work “in the style of” Monet, Lichtenstein, or Family Circus have been able to scrape that off yet.

Because of this baked-in deficiency in the final product and everyone still figuring out how best to force the AI to do our creative bidding, few serious artists have spent much time using these tools. And as all AI lacks sentience, let alone creativity, they are fundamentally just tools.

While there may be a bit of gategeeking going on with the above Colorado State Fair Chicken Littles, they’re doing it in the wrong direction. Anyone should have means and access to make art, but only those who are producing work with specific intention get to call themselves artists. Punching in little prompts in and letting the AI pop out results and then refreshing until you find something you like? That’s just a modern type of Spirograph or flipping the desktop toy over so it makes a novel mountain of multi-colored sand. (That’s no knock on either of those toys, by the way. Both can be elements of jaw-dropping true works of art. It all comes back to intentionality.)

Intentionality is indeed the trickiest part of AI-generated art at this stage for the time being, still limiting it to the realm of parlor trick. To produce something even close to an approximation of what the user is starting with in their head, they’ll have to go through a lot of trial and error tweaking and adjusting to the unknown whims of an individual AI program. And even if the AI does spit out something the user deems acceptable, this will invariably be a compromise as the path between initial vision and final product is still far more complicated and arduous than traditional mediums when you’re going for the sort of specificity and intentionality that turns a creation into art.

To prove this point, I thought it might be fun to ask a few of the more popular online AI art generators to help me anthropomorphize famous PNW landmarks. I wasn’t going to be a stickler about it. Make them cute and kawaii, make them scary creepypasta, make them as photorealistic and boring as Obama’s new presidential portrait. Just make them human.

It made the most sense to use Seattle’s Space Needle, the most iconic edifice in all of Pacific Northwest landmark-dom, as a staging ground to work out the kinks in my process. I hoped that, with enough trial and error, I’d be able to come up with a sort of universal prompt for Dall-E that spit out comparably adequate results for each of the landmarks. As a baseline benchmark, and because I wouldn’t ask a program to do something I couldn’t do myself, I took pen to notebook and cranked this out in a cool 00:01:13. It’s not the prettiest or most colorful thing I’ve ever drawn, but I wanted something for the AIs to strive toward.

The goal for this experiment

I started on the Dall-E Mini, offering the text box what I thought were some easy layups. After failing to garner anything satisfactory with “Seattle Space Needle as a person,” I switched to “…as a human.” Then I tried “anthropomorphized” as both suffix and prefix. The results weren’t what I wanted but, much like the program, I was learning.

Specifying that the Space Needle was to be a “comic strip” or “manga,” character gave me flashes of hope. Weirdly, when I asked Dall-E to render the Needle as a Simpsons character, it gave me the landmark dressed like Spongebob.

The Space Needle “... as a comic strip character” (left) and “ a Simpsons character” (right).

We weren’t getting anywhere close to what I wanted and I’d already sunk 20 minutes into experimenting, so I tried giving the landmark a task that only a human (and some primates) could do: “…shaking hands with someone.” The mutants produced told me it was time to move on to another art generator.

Two attempts at “...shaking hands with someone”

Pivoting to Midjourney’s art bot, which I operated via discord text box prompts, gave me immediately better results from an artistic standpoint. “Anthropomorphized” was starting to bear fruit, though not the variety I was after. If anything, the program had a more restrained concept of anthropomorphism than me, providing colorful renderings of the Needle where, if you squinted, you could see it as a looming villain from a Genndy Tartakovsky show. It was inherently more artistic than the rejected Quiznos commercial magazine cut-out art my last prompt had produced, but it wasn’t the art I was after.

Two Space Needles taking a rest after menacing Samurai Jack

Attempts are subverting what I thought were the program’s biases didn’t work well. “Man disguised as Seattle Space Needle” and “man wearing Seattle Space Needle costume” were total flops.

Throughout the process on Midjourney, I could see other users’ creations popping up in the chat. Taking note of the specificity of their verbiage, my next prompt attempt was “…as a cartoon character with a torso and limbs and a face.” Surely the bot would get what I was after now. I guess I was still in the realm of inconceivable fantasy to this program, because I wasn’t given any of the human features I asked for. But it had inadvertently given me something I was happy with. I asked for more and more variations and upscales on one of the bot’s quadrant offerings and, within a few minutes, I had a cute hands in pockets teen Space Needle that looked to me like a cross between a Fat Albert character and the guy from the “Paranoid Android” video.

I call him “Spacey.” We’re taking it back.

With one adequately anthropomorphized landmark down and my free trial of Midjourney nearing its conclusion, I decided to close out with a Portland landmark that would, in theory, be incredibly simple to turn into a human as it already was one: the Paul Bunyan statue. The red checkered shirt-wearing brawny boy had a face and limbs and all the human features the Needle lacked, so all I had to do was ask Midjourney to present “Portland Paul Bunyan as a real person.”

(From Wikimedia)

I got this one prompt in before Midjourney quit on me faster than any pen or tube of paint ever has. I’d used up all my allotted trial images and closed out with it failing to understand the assignment at all. Save for one on the bottom left, my Paul Bunyan options were still statues and all were far from “real” people.

Most still looking statue-ish

Reddit, some state fair judges, and digital artists already living hand to mouth may see AI generators as the future of artistry, but if my time messing around with these programs taught me anything, it’s that if you want something artistic done right, you’re still going to have to do it yourself for a long time to come.