Recent prompts: "draw the way a cat looks when it's depressed", "draw a cute raccoon using a computer", "draw a lightbulb that is trying its best".
When OpenAI came out with a new machine learning model called CLIP on January 5th, we knew we wanted to experiment with it. CLIP is sort of like the world's best AI caption writer. We thought it'd be fun to put it to the test by having it score user submitted drawings against crazy Cards Against Humanity style prompts. So, a user draws, and then CLIP assigns it a score based on how similar the model interprets that illustration to the prompt.
Since building paint.wtf this weekend, over 90,000 drawings have been submitted so far! We've seen people try to game the model by writing the prompt in text or submitting irrelevant content (but slight tweaks have been able to work around them). We've been really impressed with CLIP's resilience.
I played the prompt "draw a fire-breathing hot dog" and quickly ginned up the following entry, pictured above.
Alas, the AI was super harsh; I was ranked 58 out of 62 submissions! Waaaaah.
Looking at the other submissions, though, I can see why some of them beat me out. This one was #4 on the rankings when I last checked (the rankings change as more people enter), and it's … a pretty good fire-breathing hot-dog!
On the other hand, this next entry was ranked (when I checked) 57, so, slightly above mine …
I'm … what?
I WAS ROBBED