AI and Art kinda wondering
So why isn't there a thing where you can type like "new Taylor Swift song about eating fish; instruments: vocals, vibraphone, and bagpipes" and get a pretty convincing mp3?
I'm reasonably sure that all the tech exists?
My current half-serious theory is that the record companies wrote the AI companies to let them know that if they tried that, they'd be sued into oblivion.
AI and Art kinda wondering
@ceoln
Like this or...? https://jukebox.openai.com/
AI and Art kinda wondering
@april
Exactly! Only with the most recent improvements that have make Midjourney, ChatGPT etc so notable, easily driven by typing in a good prompt, and open to anyone willing to pay the $25/month (after 12-song free trial membership).
Seems like a missing piece of the picture...
AI and Art kinda wondering
@ceoln I would be surprised if open AI isn't working on this as we speak (yes, even on Sunday). And the generated songs will be performed by groups akin to Gorillaz.
The creator of the right AI would own all the creative cycle and profit, profit, profit...
AI and Art kinda wondering
@eperales
Could be! They had that one "jukebox" Weblog entry back in early 2020, but I don't remember anything since. If they were to release something really good that was based on training on lots of tracks from major record labels, I'd bet there would be a major copyright lawsuit or three.
But maybe that's not really what's happening, I don't know. :)
AI and Art kinda wondering
@ceoln I believe talking about copyright of AI generated art is worth of a whole new thread...
AI and Art kinda wondering
@eperales
For sure. :) I've weblogged considerably about this, and pretty much came to the conclusion that there's no telling what (for instance) the US copyright office / courts will do, because the laws and precedents are so messy! Happy to discuss more, also.
https://ceoln.wordpress.com/2022/10/31/weirdness-from-the-copyright-office/
AI and Art kinda wondering
@ceoln I suspect it’s only a matter of time, but I think humans are far more attuned to music than we are visuals. That is, our innate pattern-matching means we’re comfortable finding order in near-randomness visually, but I think it works against us in music, because anything even slightly discordant or off-rhythm stands out glaringly.
AI and Art kinda wondering
@ceoln Put another way, with images, the model doesn’t have to “understand” anything about what it is depicting, it just has to put this color next to this other color. With music, it needs to “understand” rhythm and so on.
AI and Art kinda wondering
@pwinn
Have you listened to the tracks in the OpenAI "jukebox"? I think one thing we've been discovering lately is that things can be done by generative neural nets, that we would previously have said required understanding...
AI and Art kinda wondering
@ceoln I have, and I think it supports my thesis. Every output seems even more “wrong” than even the best image generator outputs, at least to my ears.
AI and Art kinda wondering
@pwinn
Fair enough! It does seem to me to get things like rhythm and key about right, at least enough that if someone were to apply the last year's advancements to it, it could be amazing.
And that's on the harder problem of doing it at the sonic level. I wonder if anyone's tried at the (easier?) MIDI or sheet music levels lately.
AI and Art kinda wondering
@ceoln Okay, now *this* seems like what you're talking about: https://www.riffusion.com/about
AI and Art kinda wondering
@pwinn It does! Now that's pretty wild. No vocals I guess; we'll see if anyone does anything in that area. :) Maybe human voices are still too prone to Uncanny Valley effects? Hard to say !
AI and Art kinda wondering
@ceoln that's a fresh horror that I hadn't considered.