Some of the recent image-generating models have this thing where they can fill in the blank parts of images. It’s handy when you want to show them exactly how to give you more of the same. Like these animal emoji. See if you can tell which ones I gave DALL-E2 as examples, and which ones it added.
As opposed to this attempt, in which I used the same text instruction (“A grid of animal emoji”) but didn’t give DALL-E2 any starting example emoji.
Again, a few examples make a difference. Although it’s interesting that DALL-E2 can make both sets of emoji, but for some reason treats the first kind as more statistically probable when given no other instruction.
A takeaway here is that internet-trained image generators (and text generators) are trained on so much random stuff that it’s hard to get them to generate the specific kind of thing you’re after. The people making text generators will insert hidden text in the beginning of every conversation to set the stage, and it turns out a similar strategy can pay off for image generation.
The results I’m getting aren’t exactly stellar, even when I seed them with real emoji – image generators are better at vibes than at specific details, and generating small, clearly readable icons is a tough task.
Therefore I can perform a self-roast by asking DALL-E to complete my list of frequently-used emoji.
Apparently I have complex thoughts that can only be expressed using an elf with a monobrow.
I also liked “Beverages”.
And “arrows”
These “message emoji” may come in handy?
Bonus post: some more examples of extrapolated sets of emoji.
