Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

MusicGen: Simple and controllable music generation

ai.honu.io

Post reply on HN
Print view View original

191–200 of 340 posts

Re: MusicGen: Simple and controllable music generation

#191

Post by Kye »

jimbokun wrote: post #106
yantrams wrote: post #53

I'm just weirded about the fact that conversation about something AS EPIC AS THIS is so boring and rudderless here on hacker news of all places. I mean like YESTERDAY I did not have this superpower to summon something as majestic as say https://fb.watch/l4ssOD40M4/ with a simple 'A quirky and skronky Aphex twin sample that just hits you' Edits: I woke up to this news delivered from Yann Lecun himself in the morning o…

Because we like music for the fun of making it, and the shared emotional connection felt with the artist (whether it's real or not, knowing a human wrote and performed a piece allows you to imagine this connection). I don't know what the point of machine generated music is. Just destroying one of the few remaining ways for people to make a living doing something creative, I guess. The promise of automation was to hav…

Yep. The way commissioners react when I deliver the files and they hear what I made for them for the first time tells me AI has a long way to go. I'm not sure it can replace that human connection. There's plenty of solid, cheap, and sometimes even free library music out there if you just want music of some sort for a project, and no generative music I've heard comes close to it.

Re: MusicGen: Simple and controllable music generation

#192

Post by grugagag »

visarga wrote: post #156
jimbokun wrote: post #106

Earlier quoted context omitted.

Because we like music for the fun of making it, and the shared emotional connection felt with the artist (whether it's real or not, knowing a human wrote and performed a piece allows you to imagine this connection). I don't know what the point of machine generated music is. Just destroying one of the few remaining ways for people to make a living doing something creative, I guess. The promise of automation was to hav…

> I don't know what the point of machine generated music is. Exploring the latent space of human music. It's a cultural mirror.

But at the risk of aligning mirrors to other mirrors and hollowing out the essence of it. Computers have been essential to the evolution of modern music, AI won’t evolve it anywhere because it needs to mirror the human work, and without people to do that it’s a sad dead end. But I doubt people will stop learning instuments and stop making music the old way because it is too fun and meaningful to do that. But there’s a possibility it will shift in magnitude in either direction. Hope to go the way chess did and not press a button and a few faders and call it music.

Re: MusicGen: Simple and controllable music generation

#193

Post by obiefernandez »

usaar333 wrote: post #155
obiefernandez wrote: post #4

Music is already one of the most extremely devalued art forms given how oversupplied it is. Boggles the mind to think of the consequences of technology like this reaching quality levels where the differences between it and professionally produced music are imperceptible.

If anything, the consequences seem minimal because it is so oversupplied.

This comment made me realize that I should have said the consequences for me and people like me who are trying to break through new artists.

https://soundcloud.com/obie for reference

Re: MusicGen: Simple and controllable music generation

#194

Post by rubicon33 »

anlaw wrote: post #79
rubicon33 wrote: post #7

Maybe I'm just getting older but I feel like the quality of both music and film has seriously declined over the last 5-10 years. Maybe the good stuff is still out there but lost in a sea of average garbage that has surfaced to the top. Something tells me AI isn't going to rescue us either. I just sampled a bunch of these generated tracks and they immediately remind me of the average, mediocre, soul-lacking content th…

It’s not just you; pop music compositions with key changes that add complexity have all but vanished since the early 00s: https://flowingdata.com/2022/11/22/decline-of-key-changes-in... Automation tools and fine grained computed metrics have rounded off the edges of emotional experiences. See Bobby Kotick about taking the fun out of games: https://www.escapistmagazine.com/bobby-kotick-wants-to-take-... Him saying tha…

Thats honestly what it feels like. It feels like all music and film has regressed toward some boring mean. There's not enough range, emotion, and difference to find tracks that really stand out from the crowd.

Music especially just feels flat. Maybe that's just the style now, and I'm old and can't appreciate it.

Honestly, gaming is in a similar rut although not quite as bad thanks to VR.

Re: MusicGen: Simple and controllable music generation

#195

Post by yyyk »

ben_w wrote: post #49

So, CC-BY-NC licensed model weights, and they've made sure to license the training data. And some jurisdictions are saying that copyright cannot be claimed on the output of such models. Oh, to be a fly on the wall in RIAA corporate offices… Sans schadenfreude, I think this (depending on inference speed) could be perfect for dynamic content in games (including IRL games: LARP, escape rooms, table top games, etc.)

>Oh, to be a fly on the wall in RIAA corporate offices…

In all likelihood, they're ok with events. Games were never anywhere near their main revenue stream. Now the labour costs on what they're actually selling are dropping to zero. RIAA's future:

1) Use AI to fake a band.

2) Use AI to write music (maybe even lyrics). Don't really care if the AI is any good.

3) Distribute output widely, note that copyright still applies to the output.

4) Use media to generate hype (the critical step). This depends only on platform control/relations, and they have that.

5) Yea, other people could technically generate same quality dreck with AI, but it won't be (and legally can't be) exactly like the hyped dreck. Others can replicate nearly everything except the hype.

6) Since the costs are near zero just about every sale is pure profit.

Basically, since Music can be replicated, they'll sell hype and belonging to a fan group instead.

Re: MusicGen: Simple and controllable music generation

#196

Post by boredemployee »

waboremo wrote: post #175
bluefishinit wrote: post #137

Earlier quoted context omitted.

> I don't know what the point of machine generated music is. One point is that music fans can now make their own music. I think it's great that people can express themselves and it's not limited to those who put in 10k+ hours to master a single instrument. More people creating is a good thing.

The idea that creating music has had a huge technical barrier is laughable. It has not existed for ~20 years. Artists like Tinashe have learned to produce music themselves with programs like ableton, not a lick of mastering instruments or graduating from this and that art school. Just a general sense of what sounds good to you. Unlike visual art, there's no mechanical barrier either, no mastering of techniques. You c…

>> After all, "why learn how that music program works when I can just generate it?"

That's how many of my friends in the music biz are thinking right now.

Also the same applies to Code and anything that could be generated by AI. I honestly lost the joy of learning a programming language with the advent of GPT.

The future is dark.

Re: MusicGen: Simple and controllable music generation

#197

Post by emporas »

mtlmtlmtlmtl wrote: post #164
emporas wrote: post #55

I used MusicGen yesterday to create 50 songs or so. Three of them sound pretty good [1][2][3]. MusicGen is definitely the best of four models of the presentation. I used the prompts differently than the article and i think i got better results. Suppose there is way to measure cardio beats or electricity spikes on the brain, and we configure the machine to generate music to increase cardio beats, or decrease them, or…

People have tried this, they're called binaural beats, and they don't seem to work for the most part. I mean, not in the sense that you could engineer sound to invoke very specific effects in the brain consistently.

I personally have more than 10 years experience on sitting in the cold all day long, with only summer clothes on. Like 0 to 5 Celsius, with only shorts on, not even socks. I am winter swimmer as well. I do that, because i can think a lot more clear in a cold environment, it is good for the brain. Granted in Greece there is not that much cold, maybe 1 or 2 months of 0-5 Celsius.

That can be achieved by putting music on, which speeds up the heart pulse. Usually hard rock, metal, thrash metal etc. In that case, the body starts sweating a lot, not matter the temperature. I combine that, with 5 simple exercises i do all day long which are important as well.

My point is that using music, someone can be in charge of his heart pulse. But my biggest complaint always was that these metal guys, are masters of the guitar, but other kinds of music have better taste in rhythm, in melody etc. Using programs like that we can evolve it a little bit, to be more pleasurable to listen.

I know about about binaural beats, i have tried to listen to different hertz for hours on end, they don't work in my opinion. At least in my case.

Re: MusicGen: Simple and controllable music generation

#198

Post by cypress66 »

bottlepalm wrote: post #40

(and now for the rare take that isn't your typical cynical/jaded internet comment) Wow this is more than good enough to use for background music in video games, stores, commercials, etc.. You really could have super dynamic music in a video game for instance that changes based on the time of day, environment, situation, mood, etc.. all combined. Combine it with a LLM DJ and you could get some fun radio stations.

> Wow this is more than good enough to use for background music in video games, stores, commercials, etc..

I spent like 6 hours yesterday playing with this. It's really cool, but not that good yet.

I'd say it's like the original stable diffusion (without any of the finetunes and improvements). Very cool, but not 100% there yet.

Re: MusicGen: Simple and controllable music generation

#199

Post by scns »

mtlmtlmtlmtl wrote: post #164
emporas wrote: post #55

I used MusicGen yesterday to create 50 songs or so. Three of them sound pretty good [1][2][3]. MusicGen is definitely the best of four models of the presentation. I used the prompts differently than the article and i think i got better results. Suppose there is way to measure cardio beats or electricity spikes on the brain, and we configure the machine to generate music to increase cardio beats, or decrease them, or…

People have tried this, they're called binaural beats, and they don't seem to work for the most part. I mean, not in the sense that you could engineer sound to invoke very specific effects in the brain consistently.

Huberman cited a study proving the effect. Sorry, no link at hand.

Re: MusicGen: Simple and controllable music generation

#200

Post by ramoz »

cypress66 wrote: post #198
bottlepalm wrote: post #40

(and now for the rare take that isn't your typical cynical/jaded internet comment) Wow this is more than good enough to use for background music in video games, stores, commercials, etc.. You really could have super dynamic music in a video game for instance that changes based on the time of day, environment, situation, mood, etc.. all combined. Combine it with a LLM DJ and you could get some fun radio stations.

> Wow this is more than good enough to use for background music in video games, stores, commercials, etc.. I spent like 6 hours yesterday playing with this. It's really cool, but not that good yet. I'd say it's like the original stable diffusion (without any of the finetunes and improvements). Very cool, but not 100% there yet.

my first gen is good enough to be an actual song (rap beat)
Post reply on HN