I miss crappy AI art

In the last few years AI image generation has gone from something very rudimentary and nigh-on unusable for commercial purposes to the go-to option for small businesses' signs ¹ Turns out the friction of having to learn photoshop was a crucial firewall protecting society from the terrible taste of small business owners— @jdcmedlock and menus from SF to Rome.

Unfortunately I didn't ask if they would give me a 20oz cortado for $5.25.

Local flyers! So many Matcha places! Restaurant menu images! F*ck it, the whole menu! ²

As it's spread I, and everyone else, have repeatedly pointed out the same thing—this sucks. It's directly bad for menus (where showing the wrong food is worse than no food at all ³ Though Uber Eats said "screw that" and doubled down on AI photography. ) but all image generation has, in the pursuit of becoming 'better', become bad.

In lieu of six-fingered hands ⁴ "And remember, this is for posterity, so be honest. How do you feel?" the new models fall into different uncanny valleys and repeated tropes: overly busy signs lacking in clear CTAs, ornate elements that are never self-similar, the eerie feeling of everyday objects being rendered by someone who's something that's never seen them. ⁵ Somewhat reminiscent of Dürer's Rhinoceros, his 1515 woodcut based entirely off someone else's written description and brief sketch. This is worsened by the homogenization: everyone is using identical models, everyone is using simplistic prompts ("make me a poster for ____"), and so everyone's output feels the same.

A collage of AI-generated flowers from John Hartnup's recent post

The homogeneity is solvable: you can use more specific prompts and create benchmarks on creativity and leverage things like Midjourney's sref seeds. But all of this creates better simulacrums of the real, not something novel. Rather than wanting the models to get better, I find myself missing when they were worse.

DALL·E 2

While I was familiar with some of the earlier GAN-style ⁶ Generative Adversarial Network — you basically have two models, one that creates images, and one that tries to differentiate between the created images and real images. Run them against each other for long enough, and get the outputs close enough, and voila! Realistic generated images. image generators, OpenAI's 2022 release of DALL·E 2 felt like a step change: suddenly you could type in something like A painting of a fox sitting in a field at sunrise in the style of Claude Monet and make out the fox! With enough wishful thinking you might even be able to convince yourself that it was an example of impressionism.

DALL·E 1 DALL·E 2

It certainly wasn't Monet though. ⁷ I mean, just look at something 'similar' like White Frost, Sunrise from his Haystack series.Night and day. Instead it felt like something completely different, coherent from afar and yet breaking down as you zoomed in.

Grass Shadows Fur

The images felt more like a pastiche: not of art styles, but of brush strokes and colors and training images, like a ransom note of cutout newspaper letters, or Truman's paramour.

Thanks ransomizer.com Honestly, he didn't do that bad a job

This showed up in nearly all of DALL·E 2's images. ⁸ If you want more examples, this Reddit thread from launch day has a bunch of links. Take this prompt someone asked Sam Altman to generate on launch day: ⁹ Access to the model was restricted via waitlists until September of that year.

"A rabbit detective sitting on a park bench and reading a newspaper in a victorian setting" - @MikkoKuitunen3

Source from Sam's Twitter

At a first glance (and especially if you're encountering this on a smaller phone) it looked pretty good:

And he did!

But again, zooming in reveals that it's this cobbled together mess of things that look almost right from a distance, and are so deeply wrong up close. ¹⁰ To give people credit, they admitted to this too:

Center bench Rabbit hand Bench arm

Prima facie, the new models can do this better. The same prompt into ChatGPT's Image 2.5 gives a much more realistic image, replete with ornate benches, readable text, and well-shaped hands. ¹¹ Even it does try to give the gloves fingernails.

Yet it somehow feels immediately forgettable while DALL·E's image lingers. DALL·E's is a flawed image, sure, but somehow it feels more like art—surreal instead of just not real. It's doing something that modern models don't: creating something a human wouldn't.

The sound of failure

Brian Eno has a great quote about the way mediums fail being what defines them, especially after they're avoidable.

"Whatever you now find weird, ugly, uncomfortable and nasty about a new medium will surely become its signature. CD distortion, the jitteriness of digital video, the crap sound of 8-bit - all of these will be cherished and emulated as soon as they can be avoided. It’s the sound of failure: so much modern art is the sound of things going out of control, of a medium pushing to its limits and breaking apart. ¹² Shoutout to A$AP Mob's Yamborghini High music video for how it uses datamoshing. The distorted guitar sound is the sound of something too loud for the medium supposed to carry it. The blues singer with the cracked voice is the sound of an emotional cry too powerful for the throat that releases it. The excitement of grainy film, of bleached-out black and white, is the excitement of witnessing events too momentous for the medium assigned to record them."
- Brian Eno, December 19, 1995 from A Year with Swollen Appendices

Now that the hallucinatory pastiches of DALL·E can be avoided, like clockwork I've come around to cherishing them. Every time I see an AI sign for whatever "Thai Tea Slush" is, or a Summer Wagyu Fever deal I wish for an era when the best was bad, and the flaws unique.

Unfortunately it's too late—DALL·E 2 is closed source and shut down, the prompt generator and history at https://labs.openai.com have died the ignominious death of link rot, and all that's left, like vast and trunkless legs of stone amongst the level sands, are what people have saved. It's crappy.

And I miss it.

A gallery of five images I created in August 2022 to mirror DALL·E 2's watermark