Vienna
Joined August 2012
- Tweets 5,387
- Following 1,142
- Followers 86
- Likes 5
The diversity primarily comes from the variety of closeup and distance shots, I think. I love the saddle made of ruby.
0
0
0
2
This seems to be a general trend. Think about where the thing that you want to discourage commonly occurs. How it'd actually be described. And then discourage that whole situation.
1
0
0
1
Much more effective than trying to precisely and exhaustively discourage specific instances
1
0
0
1
Both images use VitB, have 4000 samples, expand 0.2, progressive growth
left image goes from diversity 1 to diversity -1
right image goes from diversity -1 to diversity 1
Asking for "Cookie Cat is a super yummy pet for your tummy who left his family behind."
1
0
0
1
one thing I'd really love (and I understand it's hard since all of this is automated) would be some sort of explanation of what defines a genre.
Alt Z is a unique enough name that I'd think y'all had an idea in that case though. Are there any notes on these thigns?
1
0
0
0
This might accomplish making it not settle into a local minimum as much. Intentionally veering off the most obvious future path to discover something else that also fits.
1
0
0
0
Novelty might amount to drawing a few things and then going with whatever shifts weights the most into a new direction (perhaps something like, the cos similarity of the proposed current gradients with the previous one? - Pick the least similar)
Not sure if that's quite right tho
1
0
0
0
On that note, I think you did a great job pointing out potential biases as you spotted them. Unfortunate that those exist but very hard to avoid, especially with such a low-curation data-set Clip was trained on. - One thing you didn't mention: The general bias of an "evil" face
1
0
0
0
I mean I know all sorts of depictions of "evil" will look like that, with distorted, asymmetric, warty faces and such, but many of those things are rooted in various stereotypes. So imo this would also fall under the "problematic" idea you put down.
0
0
0
0
Like, to me this feels very much like how it's difficult for some people to learn the difference between left and right or orientations of letters or all that
1
0
0
0
(ok those commas are confusing. I'm saying I know this is on listener habit data rather than any sound quality, but the question is how did you determine these categories like danceability or aggressiveres or whatever)
1
0
0
1
All of Zach Callison (just one album thus far open.spotify.com/artist/57YF… ) is packed into that genre and I'm guessing that's due to associated listeners, but it's WILDLY different from other stuff there so it still seems very weird. Not sure what genre that would be. But Otacore???
1
0
0
0
But anyway, I've been listening through The Sound of Everything for a couple weeks now. Currently at G. Discovering a lot of stuff there's no way I'd ever have found it otherwise in the process. So thanks for that! (Also one of the "Didgeridoo" tracks is 100% organ 0% Didgeridoo)
1
0
0
0
Now it sounds like your approach to podcasting is much less meandery, so the consistent one hour eps would probably be a stretch. But I think half an hour is eventually actually quite realistic. (Early on maybe not: The context builds over time, giving you more to talk about)
1
0
0
0
There are also small inconsistencies between Desktop and Mobile Discord. Some mod actions are practically impossible on mobile. Meanwhile, you can't upload multiple things in a single message on Desktop.
1
0
0
1
It should also be possible to discover emote reactions via search.
A proper built-in mod hub (where you can keep track of people to watch) would be good (instead of relying on bots that keep going down over server issues etc.)
Perhaps it should be possible for mods to edit posts.
0
0
0
1