←back to thread

Kagi Update: AI Image Filter for Search Results

(help.kagi.com)
270 points lkellar | 4 comments | 17 Oct 24 20:02 UTC | HN request time: 0.006s | source
Show context
nox101 ◴[17 Oct 24 22:47 UTC] No.41874589[source]▶
So that's nice for people that don't want AI images. In the other hand, I actually want the opposite. I want AI images ranked higher. They're more likely to be usable (CC-.., PD, etc)
replies(1): >>41874632 #
1. freediver ◴[17 Oct 24 22:53 UTC] No.41874632[source]▶
Would you want to search for them though? Or ability to generate them yourself easily?
replies(1): >>41875485 #
2. nox101 ◴[18 Oct 24 01:10 UTC] No.41875485[source]▶
searching for the often leads to the prompts I need to get a similar image from which I can iterate from. So yes, I want to search for them
replies(1): >>41876555 #
3. gandalfgreybeer ◴[18 Oct 24 05:22 UTC] No.41876555[source]▶
>searching for the often leads to the prompts I need to get a similar image from which I can iterate from.

how crazy are these images that you can't start with a prompt of your own?

replies(1): >>41876761 #
4. nox101 ◴[18 Oct 24 06:14 UTC] No.41876761{3}[source]▶
at least in the image generators that I've used it takes quite a bit of experimentation to get both positive and negative prompts that produce both the result you want and a quality one as well.

Part of it, if I understand correctly, is that the data has no actual concept of grammar and meaning. So if you say "a table with 6 legs" it doesn't see that as a full concept, it sees words or segments of words, so "legs" can end up being legs of people (it might make a table with human or animal legs) and similarly it just might insert random legs in the scene because it has no understanding of the description.

So, people find ways to coerce / influence it to get to the right place.

It's also important to know what models/loras/etc were used as not all data sets generate good images for whatever your topic is.