Imagine someone gives you a new edition of a beloved picture book, illustrated with thousands of newly discovered illustrations by the original artist. As you look through it, you notice that a lot of the new images seem strange and off-putting. Or they’re just plain confusing.
“Don’t worry,” your friend said. “What’s great is that there is such images that if you keep flipping, you Ultimately get something you want.
This would not be a good gift! Yet this is the value proposition of the recent Google Labs project Infinite Wonderland experience.
This online project is based on Lewis Carroll’s idea Alice’s Adventures in Wonderland (1865), and the famous 42 pictures by the Victorian illustrator John Tenniel (1820-1914) from its first edition. Google Labs used AI to clone Tenniel’s style, so that if you click on any line of Carroll’s text, you generate a whole new AI-powered image that looks like one of his original illustrations.
To go further into the present, four contemporary illustrators also lend their style to Infinite WonderlandThe experience of: Eric Hu, Shawna, Erik CarterAnd Haruko Hayakawa.
Today, at this particular stage of the AI hype, a project like Infinite Wonderland He’s unlikely to impress anyone on his own. (It’s actually meant to show off a technique called StyleDrop, which allows Google’s AI to imitate any artist’s style using just a few source examples, the company complaints).
On the other hand, the fact that Infinite Wonderland often produces sloppy or disturbing images. This is unlikely to be all that surprising, although I really don’t see how this isn’t considered a crippling bug in this case, given that the proposal is a children’s storybook. (Each image features the amusing disclaimer: “AI output may sometimes be offensive, inaccurate, or overly surreal. If so, please regenerate.”)
Just consider one of Tenniel’s most famous images, showing the moment Alice meets the caterpillar:
Illustration of Alice’s Adventures in Wonderlandby John Tenniel.
If you click on the corresponding lines of Infinite Wonderlandyou will probably get something like this:
Illustration of Infinitive Wonderland in the style of John Tenniel.
This is not an improvement.
What about other living illustrators? Here’s an AI rendering of the same scene in the style of Erik Carter, who specializes in pixel art-themed graphics. For some reason, it shows the caterpillar on the ground looking at Alice, instead of the other way around, like in the text. And the caterpillar’s expression is enthusiastic, instead of “languid and sleepy,” according to Carroll. And he doesn’t have a hookah.
These are fundamental errors that any illustrator would have corrected.
Illustration of Wonderland infinitive in the style of Erik Carter.
Refresh for an illustration of the same scene in the style of Shawna X. Now, for some reason, our caterpillar lights a pipe for Alice.
Now that’s great. Unintentionally awesome, but awesome.
Illustration of Infinitive Wonderland in the style of Shawna X.
Any artist creating images for a book would have hundreds of ideas and then narrow them down to a few that would be most appropriate. In practice, Infinite Wonderland simply outsources this work to the reader, who must click and click again until they get something they like, much like Ikea makes furniture cheaper by outsourcing its assembly to you. Except that the Ikea table is finished when it’s finished, and Infinite Wonderland is unstable by design, constantly moving around like a bad dream.
Again, this is all bad, but not even really bad when it comes to enterprise AI initiatives. Infinite Wonderland was neither good enough nor bad enough to garner much attention when it came out a few months ago (aside from a credulous article in Fast business). The reason I wanted to say something about it is the subject matter.
Obviously the only reason to use Alice in Wonderland is due to its proto-surrealism and its associations with wonder and breathtaking experiences. AI art is often associated with surreal juxtapositions (which is why DALL-E is named after Dalí). But it turns out there’s something about the specific AI procedure Google Labs uses that’s particularly bad for surrealism.
Take this illustration by Tenniel, which shows what happens to Alice when she eats the cake that makes her grow taller, and her body expands “like the largest telescope that ever existed.”
Here is Tenniel’s illustration:
Illustration of Alice’s Adventures in Wonderlandby John Tenniel.
And then here’s how Infinite Wonderland performs the same scene in the style of Tenniel:
Illustration of Wonderland infinitive in the style of John Tenniel.
Stupidly, he focuses on the idea of Alice looking at her own feet, then interprets that as Alice looking at a pair of shoes. This is just a truncated reading of the text.
I suppose you can attribute this to the fact that the AI failed to determine what is the most interesting element of the passage to illustrate. Well, what about another passage? This is Tenniel’s original depiction of Alice playing a game of croquet, where the mallets are flamingos and the croquet balls are curled up hedgehogs. She notices that the bird keeps turning to look at her face every time she tries to swing.
Illustration of Alice’s Adventures in Wonderlandby John Tenniel.
If you locate the corresponding passage and click on it, Infinite Wonderland really, really doesn’t want to give you anything close to that fantasy scene, no matter what style you try.
Here are examples in the styles of the 5 artists of the project:
Illustration of Wonderland infinitive in the style of Eric Hu.
Illustration of Wonderland infinitive in the style of Shawna X.
Illustration of Wonderland infinitive in the style of Erik Carter.
Illustration of Wonderland infinitive in the style of Haruko Hayakawa.
The results are not only not surrealists, they are anti-surrealists. They took Lewis Carroll’s animated text and rendered it in the most boring form possible, apparently because Google’s AI insists on interpreting “playing croquet with a flamingo” as “playing croquet next to a pink flamingo.
Which makes sense, from what I understand. AI works by giving you the most likely response to a request. The whole charm of the idea of croquet based on flamingos lies in the fact that it is an unexpected image, but Infinite WonderlandCarroll’s logic pushes in the other direction. What you get is a very sophisticated technical procedure for de-surrealizing a story.
It’s not like you can’t use AI to create surreal things: people do it all the time. But this requires at least some reflection on the real meaning of specific passages, and Infinite Wonderland is so crazy that I can’t escape the feeling that it’s made by people who don’t think much about what artists do and who don’t look at works of art with much interest. It seems to me that the whole way of thinking that Google Labs is selling involves thinking less about images.
Instead of using Alice’s adventure as a magical portal to the imagination, Infinite Wonderland fills it with mud. It’s a thought I find dark enough that I end with something funny. Here are only 10 successive attempts to use this state of the art, energy consuming Artificial intelligence by one of the largest and most powerful technology companies in the world to do the same work that John Tenniel did with pencil and woodcut technology in 1865.
Here is the original (from the opening of Chapter VII: “The Queen’s Croquet-Ground”):
Illustration of Alice’s Adventures in Wonderlandby John Tenniel.
And now here is the future of art!
