I was asked by the Riffusion team to spread the word about their public release of their new model. http://riffusion.com Sound quality works well on some genres but not so great on others. Live their UI/UX and lyrics gen. Its still in development just like a lot of these tools but check it out!
I advise for suno. But have been helping Riffusion with their gen AI platform. Just like suno but their own algo, model, and approach to gen AI.
Was actually going to try this and one of the open source models that dropped recently, Yue I think it's called. Any transparency on Riffusion's training datasets or will we just be regurgitating derivatives of other musicians stolen work like all the other services?
YuE used open internet datasets like all of these companies. Cool thing is YuE can run locally (on pc) and easy to install there if you get https://pinokio.computer/ they just updated it yesterday with audio upload features. https://x.com/cocktailpeanut/status/1886456237883048107 the other cool thing with the fine tuning tech that deepseek figured out, and how things have been going in dataset training. In no time any of these music models will be able to be trained on non open internet data but still will be able to recreate any style of song. Suno went from v3 to v4 models with way less data. So yes you will be able to re-create a pop country song that sounds like something taylor swift would have made but her data will never be in the dataset. Also Good Loras will be a thing very soon in generative music so we can have our own custom fine tuned model pull from.
"Stolen" is a strong word with how AI datasets work. They just understand all the data they are given then when they are called with prompts and weights temperatures they will generate on what they understand about the music inside of the data not actual audio. you can prompt any AI tool, GPT, Suno, MidJourney to reflect the information in the dataset. Or you can use these as tools to not be a mirror and actually generate something unique. Blend elements from multiple genres, or get weird with the data. If you are a guitar player and you were influenced by led zepplin you essentially are pulling from that dataset when you perform but you dont have the wav file of ledzepplin in your brain.
if you want to get into training your own datasets to use with music gen. This is a good google colab that uses stable audio. You can make your own lora with it and run inside of stable audio: https://colab.research.google.com/drive/14-uBFtWMELsRFc9jx-qMX5OkfKdB0vnf
