Train to be a perfect loop

[cacosta88]

Train a separate model to perfect the 30-second loop, ensuring it seamlessly folds/loops onto itself. This will make the output easier to use for musicians/producers, even in its current state with all of its limitations.

[FurkanGozukara]

this would be amazing

[tmc]

What would need to go into this?

[chavinlo]

https://github.com/chavinlo/musicgen_trainer done

[cacosta88]

@chavinlo, it seems to work for the example you provided: Overfit -> Generate -> Decode: https://voca.ro/1f6ru5ieejJY. It does sound like a perfect loop. However, for the Bossa Nova example, it doesn't sound like a 30-second sound that can be seamlessly repeated.

For now, I am actually doing it manually on Ableton, but it takes a long time. I imagine one could train a model using a bunch of perfect loops from https://splice.com/. The target training output would be the original loop or the last few seconds, and the training input would be the original loop with its last second or two chopped off. Then for inference, you would input the imperfect Audiocraft loop, and hopefully, get a perfect loop out of it.