Intrigued by J Greely’s work with pinups, I’ve been experimenting with AI art generators. Rather than struggle with SwarmUI, I tried some of the many free online toys. The results were interesting enough to warrant further exploration, and I eventually ended up at NightCafé. I’ve been seeing what the various models can do and what their limitations are.
As far as capabilities go, they can imitate almost any style to some degree. AI “photographs” are convincing as long as you don’t count the fingers, and sometimes the models get those right, too — one more reason not to believe anything you see online.
However, despite the millions and billions of parameters the models are based on, there are serious lacunae in their knowledge.
I asked a number of different models to show me Dagwood Bumstead and Sailor Moon as they should appear, given their caloric intake. All of them knew Usagi, but none had a clue who Dagwood was. This was unexpected, since Blondie has been in the newspapers since 1920, while Sailor Moon was first published in 1991. I asked for a picture of G.K. Chesterton and George Bernard Shaw arguing at an outdoor café. Only one (Nano Banana (via Google Gemini1), I think) came up with an image that resembled the pair at all (top); the rest just showed me two generic greybeards at a table.
It’s easy — too easy — to make striking images. One-word prompts can produce dramatic, unexpected pictures. However, after a while they all start to look the same. Looking at the steady flood of new images at NightCafé, I see the same faces, the same streets, the same landscapes, the same cheap surrealism, the same compositions over and over and over. There may be a vast number of permutations of the elements possible, but despite their millions of parameters, all the models ultimately seem limited and formulaic.
I’ve tried to come up with some non-clichéd pictures, but it’s not easy. No matter how carefully I write the prompts, Seedream, Flux, HiDream, Ideogram or whatever will misinterpret half of what I write and ignore the rest. I don’t so much “prompt” the models as try to outwit them. Usually I eventually get an acceptable image, but it’s never what I envisioned. I suppose I’ll get better at this with practice, but it might be easier to just learn to draw.
So, are computer-generated images art? Nah, they’re wallpaper.
The morbidly curious can see more of my “art” here.
Notes
- I mention this with reservations: Google is evil, as we all know. You might want to take steps to minimize its spying.
The problem with characters is the tagging; between sites like Danbooru and DeviantArt, Sailor Moon has orders of magnitude more tagged images than Dagwood. This is why there’s a thriving trade on Patreon and other sites where people make money creating custom LoRA models for specific characters (and sites like Civitai where people do it for free).
Related, Today I Learned that Modern Dagwood carries a smartphone and plays games on it. The writing is… not good. I feel no desire to start reading newspaper comic strips again…
-j
Did you ask for the “musical combat” to be refereed by Tony Levin, or was that just the AI being accidentally intelligent?
That was a surprise, as was the piper’s mohawk.