9 comments

[ 2.8 ms ] story [ 38.9 ms ] thread
The recomendation algorithm is the most interesting part here - with your feed of AI images depending on your swipes. A future of infinite content.
Yes, the algorithm :p

Actually joking aside, if there are infinite results that are built up by preference it would be interesting to see what my "type" actually was. Particularly given o dont feel I have a "type".

Probably won't explore it, but I have to admit it could be interesting.

This felt like someone uploaded their ai pictures folder and put it behind a Tinder-like photo browser.

What is happening behind the scenes?

I swiped left of every super hero looking, yet they still kept coming. I swiped left on almost every anime one, and a lot kept coming. There was a random cat and some sand. I also saw several duplicates. I also swiped left on a bunch that looked like a plus sized Billie Eilish, yet they still kept coming.

I expected a lot of randomness at the start, but then more and more right swipes as time went on. This wasn’t the case. It seemed pretty random, in terms of my tastes, and seemed heavily influenced by what the author was into, and it wouldn’t let it go.

Interesting idea though. I just wasn’t feeling like it was using a recommendation algo based on my swipes. I also wasn’t entirely clear on the play button. I did press that on the first image to see what it did, and it just changed the image. Is that why it kept showing me more stuff like that, even though I swiped left?

So it's using vector embeddings - for each image. As you swipe, you get an aggregate preferences vector - which is more heavily influenced by downloads and plays. There is a time decay on the preference vector (older likes mattering less). Still tuning things for the optimal quality vs. your desires tradeoff (like purely optimising for your preference vector would lead to possibly crap images)

But maybe I can adjust a little more to preferences

What does the play button do? Perhaps an explanation of the interface might help.
so the play button creates variations of the image - it goes through a language model and the prompt is adjusted heavily inspired by the current prompt - that and the current image are used to create the next image

You can see the current prompt of an image by clicking the edit button