How to help?
Are you looking for smarter losses?
More training data? Maybe h3 minimax can regurgitate audio for training?
Could we overfit to a given song?
yeah we kinda hit a ceiling with synthetic data and real audio requires finetuning the LM, if you experiment with getting vocals into the model, we have solved style for the most part
https://btdig.com/75d765537660d7a9c0461d3992d0e50275bfde3f/mp3
https://btdig.com/ec5f0bb5bbfacd51dedaf16a3ec1c06f2a62c198/mp3-indian
https://grok.com/share/bGVnYWN5_ae71818f-2b7b-4e14-823d-f449e6c15ac1
I'm not sure exactly if this is a distraction -
https://raw.githubusercontent.com/ServeurpersoCom/minimaxmusic.cpp/master/docs/ARCHITECTURE.md
i push some code for v5 - https://huggingface.co/johndpope/open-rvq-encoder-minimax-music3/tree/main
it has wandb - i highly recommend -
Project: https://wandb.ai/snoozie/open-rvq-minimax-music3-v5
Cover/listen run: https://wandb.ai/snoozie/open-rvq-minimax-music3-v5/runs/f5phji1d
UPDATE -
wandb no longer recommended - they updated to a 60/mth paid model.