What data was used to train riffusion? about riffusion

[oxysoft]

@scf4 Public figures working in ML like the authors of Riffusion should stick to royalty free datasets. The strategy for AI music will be to make fine-tuning really easy and available to all, easy enough for a 12 y/o to do it comfortably just by following a few easy steps. 12 y/o kids don't care about copyrights, it will be limewire all over again and this time the music industry will be utterly powerless.

[karen-pal]

I've opened a similiar issue on the app's repository: riffusion/riffusion-app#12

Dataset info is very important and will open doors for creators everywhere.

[enn-nafnlaus]

I understand why they don't disclose it, though - to them, it just invites possible headaches.

[hmartiro]

Additional fine tuning and data information has been added to the model-card. This was trained using approaches similar to hugging face examples, but fine-tuning can be achieved with very small datasets using a dreambooth approach. Will look to add tools to make fine tuning on your own data simpler.

[happypepper]

@scf4 Public figures working in ML like the authors of Riffusion should stick to royalty free datasets. The strategy for AI music will be to make fine-tuning really easy and available to all, easy enough for a 12 y/o to do it comfortably just by following a few easy steps. 12 y/o kids don't care about copyrights, it will be limewire all over again and this time the music industry will be utterly powerless.

If you check out OpenAI's jukebox project, they've clearly trained with copyrighted music and don't really care.

[breadbrowser]

@scf4 Public figures working in ML like the authors of Riffusion should stick to royalty free datasets. The strategy for AI music will be to make fine-tuning really easy and available to all, easy enough for a 12 y/o to do it comfortably just by following a few easy steps. 12 y/o kids don't care about copyrights, it will be limewire all over again and this time the music industry will be utterly powerless.

If you check out OpenAI's jukebox project, they've clearly trained with copyrighted music and don't really care.

OpenAI can get rights to them.

[jbmaxwell]

I was also curious about the initial training. I'm not so bothered about specifics of the dataset (though general info on size would help) but more about training params (lr, bs, etc.) and duration (epochs/steps).