Restructure?

[gurjeetjagwani]

I am proposing a bit of a severe restructure of the repository. I believe the way we have currently setup everything is not accurate to how these models are now viewed in the wider literature. See for eg: https://arxiv.org/abs/2303.08797 , https://arxiv.org/abs/2509.24531, or even https://arxiv.org/abs/2605.16486.

I believe we should have a structure resembling something like this:

flowfusion
-> loss.py (score-matching, flow-matching, schrodinger-bridge matching, etc)
-> arch.py (MLPs, CNNs, VAEs, etc)
-> trace.py (Hutchinson, Xtrace, Hutch++, StAD, etc)
-> pf_ode.py (SDEs, ODEs, Bridges, Symplectic-ODEs, etc)

There would obviously be more things to be included here and we would have to be careful with backwards compatibility, etc. But I believe a structure like this is more clear for users, modular and enables us to do more with our package.

I would welcome a discussion on this: @stevet40 @hiranyapeiris

Also users of the package, so: @ixkael @AlexBrown6291 @joygong , if you have any opinions about this, please let us know.

I am happy to move this as a long term goal, but we should be thinking about this if we are to submit a JOSS paper, and create a package that is useful for the wider community.

[gurjeetjagwani]

Would also help with #5 #13 #16 #17

[stevet40]

I agree in principal with this. Would be good to restructure things to better exploit the commonalities. We can sit down at some point and figure out exactly what functions should go where, etc.

100% backwards compatibility may be impossible. As a stop-gap we might be able to set up some wrapper classes that reproduce the behaviour of the existing code structure. I'd say I'm personally not too worried about this though. Another alternative would be to provide a script that unpacks an old flowfusion.flow or flowfusion.diffusion model into the new classes.