3 ms·

I have been a musician at this point for 35+ years. The quote from the Suno CEO is beyond absurd. I have spent quite a number of hours with MusicLM in AI test

I have been a musician at this point for 35+ years. The quote from the Suno CEO is beyond absurd. I have spent quite a number of hours with MusicLM in AI test kitchen. To me, that is actually a new form of music synthesis. Synthesis in the context of FM, subtractive, etc. I suspect MusicLM had some tracks removed from the training data because it use to make absolutely wildly creative psytrance clips and now it just doesn't. Suno on the other hand to me has basically been a complete joke. The training data is just not wide enough to do anything interesting and new. I think the only way to really use these tools creatively would be to produce your own trained model with AudioLM. The creative use then is in the musical output and not selling the model as yet another SaaS for $x.99 a month. Of course, you could use Suno to create loops and sample material. The way midjourney is the greatest thing possible for digital collage in Krita/Photoshop. GenAI art though on its own be it midjourney, stable diffusion, Suno, MusicLM, whatever already jumped the shark for me a year ago. The problem ultimately is the output is just so very limited.

chaosprint 2y ago
Technically, I feel that the current diffusion angle always has to fight against sound quality. But maybe this angle can find its own "Retina" moment, and Suno's v4 is closer. But the hard flaw (almost unsolvable) of this angle is editing. I have been exploring methods based on DRL( https://github.com/chaosprint/RaveForce https://github.com/chaosprint/RaveForce), but I think there is still a long way to go.
gedy 2y ago
> Suno on the other hand to me has basically been a complete joke. The training data is just not wide enough to do anything interesting and new. Hard disagree, Suno is terrific, but likely depends on what you define as "new". Of course AI models can't explore outside the data they are trained on, but they can effectively generate within the parameter space, including areas unexplored directly by the content it was trained on. Have done some really interesting mashups of styles that people aren't doing and the results are great imho. If you are trying to make it do very specific things like you would another instrument, it's not the right tool.
- teucris 2y ago
  I would love to hear them. Would you share a link or two?

chaosprint 2y ago

Technically, I feel that the current diffusion angle always has to fight against sound quality. But maybe this angle can find its own "Retina" moment, and Suno's v4 is closer. But the hard flaw (almost unsolvable) of this angle is editing. I have been exploring methods based on DRL( https://github.com/chaosprint/RaveForce https://github.com/chaosprint/RaveForce), but I think there is still a long way to go.

gedy 2y ago

> Suno on the other hand to me has basically been a complete joke. The training data is just not wide enough to do anything interesting and new. Hard disagree, Suno is terrific, but likely depends on what you define as "new". Of course AI models can't explore outside the data they are trained on, but they can effectively generate within the parameter space, including areas unexplored directly by the content it was trained on. Have done some really interesting mashups of styles that people aren't doing and the results are great imho. If you are trying to make it do very specific things like you would another instrument, it's not the right tool.

teucris 2y ago
I would love to hear them. Would you share a link or two?