Hacker News Reader

Comment by devin | Hacker News Reader

I listened to like a dozen of the examples on riffusion and they all sound like junk. It's kind of like the hands being messed up on generated images, except the "messed up hands" stream through every part of each track.

View on HN

Come on... we are very early in this new world. It's always funny the lack of imagination people have with this stuff. Have you not been observing the rapid progress of LLMs? This app was not possible a few years ago. What will these products look like a few years from now? What commercials on the Super Bowl will be using AI generated music in the background? Will the NFL Network be using these AI sounds during their transitions?

I mean, yeah most of these are lame. I found one or two that actually sounded good. But they are short clips. It's probably going to take awhile for AI to generate an entire 4-minute track cohesively.

I understand your sentiment but I have to agree with the parent. Sure, probably in a couple of years it will get better. But for now, it simply isn't good. Even generative music from the pre(current)-AI era sounded much better, even though it was quite boring.

Really, I am interested, I'm looking forward to new developments, but at the same time I have the right to call a spade a spade, to be euphemistic.

...so? The parent didn't say music generation AI will never be good. They were just commenting on one single thing, which is Riffusion.

I have no experience in this area, but I feel like the ideas are the hard part in music. These are rough drafts and some of them have promise if they can be refined in an iterative fashion with the user at the helm.

You can generate very interesting music simply by working with MIDI as opposed to sampled audio (slashing complexity by orders of magnitude!) and starting from a good model architecture. Daniel D. Johnson's (formerly known as Hexahedria, hired by Google Brain) model biaxial-rnn-music-composition is from 2015, requires very few resources for training or inference, and still delivers compelling, SOTA-or-close results wrt. improvising ("noodling") classical piano. Github https://github.com/danieldjohnson/biaxial-rnn-music-composit... , you may also want to check out user kpister's recent port to Python 3.x and aesara: https://github.com/kpister/biaxial-rnn-music-composition (Hat tip: https://news.ycombinator.com/item?id=30328593 )

Reply on news.ycombinator.com