Monday, 17 October 2022

AI Can Produce Visuals We Can’t Even Imagine, So Maybe We Should Just Enjoy It

NAB

Would you want to visualize your dreams? Might it be fun to try? There’s an AI for that — dozens of them, in fact — and even if catching those internal flights of fancy doesn’t capture your imagination, turning an AI lose on the surreal, the subjective, and the weird could be the best thing to happen to art since… well, ever.

article here

That’s the positive spin on the human art versus synthetic media debate that erupted into the open this year following OpenAI’s release of DALL-E 2.

It and other text-to-image generation engines are devilishly easy to use — which perhaps explains the angst of some artists wishing to keep their work shrouded in mystery.

The text used to produce AI images with algorithmic tools like Midjourney, DALL-E 2 or Craiyon is called a prompt, and there is a school of thought believing that there could be a role in the future for “prompters” with creative input into everything from script development to music composition.

Some worry these new AI tools might threaten the livelihoods of artists, provide new and relatively easy ways to generate propaganda and deepfakes, and perpetuate biases.

However, Jason Scott, an archivist at the Internet Archive and a prolific explorer of AI art programs, as well as being a traditional artist himself, says he is “no more scared of this than I am of the fill tool” — a reference to the feature in computer paint programs that allows users to flood a defined space with specific colors or patterns.

In a conversation at The Atlantic Festival with The Atlantic executive editor Adrienne LaFrance, Scott discussed his quest to understand how these programs “see.”

He called them “toys” and “parlor game[s],” and did a live demonstration of DALL-E 2, testing prompts such as “the moment the dinosaurs went extinct illustrated in Art Nouveau style,” or “Chewbacca on the cover of The Atlantic magazine in the style of a Renaissance painting.”

Scott isn’t naive about the greater issues at play — “Everything has a potential to be used as a weapon” — but at least for a moment, he showed us that the tech need not be apocalyptic.

Rather than limit AI tools to replicating, albeit superfast, ideas of things we already know (prompt: “tables and a cat in a room in Picasso style”), we should explore further into the realm of things that seem improbable. Things that perhaps humans are simply not capable of visualizing because our language (the words which dictate how we order and understand the world if you’re in a Wittgensteinian frame of mind) is inadequate.

But not so the intelligent machine.

“The weirder and more creative you get with this toy, the more fun it gets,” says Scott. “I see a future where you’ll be able to say, “Could I read a book from the 1930s where it’s got a happy ending and it takes place in Boston?” Or, “Can I have something where they fell in love but they’re not in love at the end?”

He wants to push the “toy” to see what happens when prompted with juxtapositions that — were you to script them into a film — would be unrealizable or nonsensical.

For instance, here is Scott’s “lion using a laptop in the style of an old tapestry.”

This is “Santa Claus riding a motorcycle in the style of 1970s Kodachrome.”

This is “Godzilla at the signing of the Declaration of Independence.”

These are “bears doing podcasts.”

This is “GoPro footage of the D-Day landing.”

Scott wants to understand what the AI systems are seeing? “It’s so easy as a parlor game to say, ‘Draw a cellphone as if it was done as a Greco-Roman statue.’ But what about doing a bittersweet sky, or trying to draw a concerned highway? What does it see?”

It is interlocuter LaFrance who brings up the idea of being able to visualize dreams.

“What other sorts of things — fiction comes to mind — can we imagine but don’t normally get to visualize?” she asks.

The human mind boggles, but AI has only just gotten started.

 


No comments:

Post a Comment