Fixing mixture vs conditioning the prior with mixture

[biendltb]

Hi there,

Firstly, thank you for your excellent work on this project.

I noticed that in your approach, partial generation is achieved by fixing known sources while generating the target sources. This method allows the model to leverage relationships among the sources to perform its task. Out of curiosity, I wonder if you've considered treating source separation similarly to partial generation. Specifically, by stacking the mixture together with the stems and keeping it fixed, the model could generate the sources based on the relationship between the mixture and the sources.

I'd love to hear your thoughts on this approach.

Thank you once again for your efforts!