LoRAs possible for fast and lightweight fine-tuning?

[oxysoft]

Just wondering if it will be feasible? Obviously the base model probably wasn't trained with any copyrighted music and probably won't recognize any band name I throw at it, so it will be very valuable. I could already foresee a huge library of LoRAs much like Stable Diffusion for every band imaginable - Pink Floyd crossed with Aphex Twin, anyone?

[chavinlo]

you first have to figure out how to train

[FurkanGozukara]

I wonder if we could somehow generate more than 30 seconds

this model amazing

[0xdevalias]

I wonder if we could somehow generate more than 30 seconds

See:

• https://github.com/facebookresearch/audiocraft/issues/8#issuecomment-1585740081
• https://github.com/facebookresearch/audiocraft/issues/36#issuecomment-1586236702

[0xdevalias]

you first have to figure out how to train

• https://github.com/facebookresearch/audiocraft#will-the-training-code-be-released\

  • Yes. We will soon release the training code for MusicGen and EnCodec.

[adamfils2]

Any updates on this?