AI Graphics and Crossword Clues

How robot eyes would've done it.

As I said earlier, getting images for the Ubercross G’s visual clues wasn’t always easy…or free. It was complicated enough that I wondered—Hey, there’s a lot of noise about AI-generated content these days. What if I got that to work for me?

Short answer: it would have been sometimes interesting, but often frustrating, and definitely would have taken way too long.

The DALL·E interface at labs.openai.com is commanding a lot of attention, all right—too much for its own good. At just about any time I tested it, there were multiple time-outs and “Sorry, we’re overloaded” messages. If I’d tried to produce all 570 of these pictures with it on a deadline…well, you can imagine the stress. It wouldn’t be free, either, though the $75 that I’d pay for that many images was a bit less than what I invested in royalty-free services.

But with patience and perseverance, I was able to see what a few samples would look like. Overall, I was glad to have saved my money.

This is the image I used for Abed from Community, followed by the best DALL-E result for “Abed from Community looking directly at us.” Clearly DALL-E either has no authority to reproduce fictional characters, or it has no idea who Abed is.

This image of Muhammad Ali is a piece of a famous shot, one of the images I fretted most about my authority to use. I decided it was altered enough to be worth the risk, but could DALL·E get me a safer one?

No. Apparently that would be against content policy. I’m not entirely sure what the issue is—a post-game victory pose isn’t violent in and of itself, even if it suggests violence has recently taken place.

DALL·E could, however, get me a picture of Thanos posing triumphantly with boxing gloves in a boxing ring…

…even if its concept of Thanos was basically “an older-looking muscular dude, and purple’s involved somehow.”

So okay, DALL-E just isn’t great at rendering people with proper names. But there were also dictionary words I needed to clue. What about an anorak?

Not bad! I specified most of the elements here: my request was “black man in yellow anorak in the rain.” I was mindful of likely color contrast. The result is a much more dynamic image than what I got with conventional methods. DALL·E wins this one.

Here’s the continent of Africa, rendered as a black outline and in DALL·E’s “impressionist” style:

Both are clearly recognizable, both get the job done. I now think the DALL·E version is a little too show-offy, but that is kind of what I asked for. Let’s call this a tie.

With “arms akimbo,” I experimented with combining inputs, asking DALL·E to show me a woman in a green anorak with arms akimbo. She’d have to make a good impression to beat this confident-looking dude:

The best of them is not so full of personality as the icon above. DALL·E seems a little underconfident about its ability to render women’s faces. Considering how “Abed” looks, I’d say that underconfidence is probably justified. This is not too bad:

But it must be said, two of the other entries didn’t seem to get what “arms akimbo” meant:

And don’t look too closely at those hands.