DALL·E 2 and images from nowhere

One of the example prompts on OpenAI’s DALL·E 2 page is an astronaut riding a horse, in a photorealistic style, and the picture next to it looks like a still from a movie I’d watch. I’ve been scrolling through the samples people have been posting since it was announced last Wednesday, and my reaction has gone from “that’s a cool demo” to “huh” to mild unease in about that order.

The first DALL·E, in January last year, made small, smudgy 256-pixel images that were impressive mostly because they existed at all. This one makes 1024 by 1024 images, four times the resolution, and it can do two other things that seem more important to me than the headline. It can edit, so you can take an existing photo, mark a region, and describe what should be there instead (put a flamingo in this pool), and it matches the lighting and adds the reflection in the water. And it can do variations, take an image and give you more like it. The way OpenAI describes it, the model is built on CLIP, their model from last year that learned to match images and captions, plus a diffusion model that starts from noise and turns it into an image a step at a time. I understand diffusion about as well as I understood backprop before I did it by hand, which is to say I can explain the picture but not the details.

Almost nobody outside OpenAI can use it yet. It’s a research preview with a waitlist, and I’m on it, along with what I assume is everyone I know. Sam Altman has been taking prompt requests on Twitter and posting the results, which is a fun way to run a demo and also means every image anyone has seen was picked by someone at OpenAI. People have pointed out the weak spots in the samples that are out there. Text comes out as nonsense letters, and if you ask for a red cube on top of a blue cube you might get it the other way around. My guess is those get fixed in a version or two, but that’s a guess.

OpenAI is doing the same staged thing it did with GPT-2 three years ago, slowly and with rules. They say they removed the most explicit and violent images from the training data, that it won’t make realistic faces of real people, and that they’re watching how the people in the preview use it. Back then I thought the staged release mostly wouldn’t stop anyone, and I feel the same way now, though I’m a bit more sympathetic, since a picture of a real politician doing something they never did is a more obvious harm than a fake news story was.

What I actually want to know, in no particular order:

What’s in the training data? Hundreds of millions of images with captions, scraped from the internet, which means a lot of them were made by people who make images for a living. When it does “in the style of” a living illustrator, that illustrator didn’t sign up for it, and I’d be surprised if this doesn’t end up in a courtroom at some point.

Who’s it for? The obvious answer is stock photos and the kind of illustration that goes at the top of a blog post (I’ve spent more time than I’d like to admit picking header images, so, I see the appeal). The less obvious answer is everyone who needs to make a mockup, a storyboard, a concept sketch, which is a lot of people and mostly not artists.

How will I know what’s real? We spent years learning that a photo can be edited. We haven’t really learned that a photo can be made up from a sentence, and the edit feature, which keeps the rest of a real photo and changes only the part you mark, seems like the harder one to spot.

Ten years ago I was doing Andrew Ng’s homework on recognizing handwritten digits, and that felt like magic. Going the other way, from words to a picture nobody drew, is not something I expected to see this soon. I’m not sure what to do with it except get off the waitlist and see how it handles a request it wasn’t cherry-picked for.