have fun!!! 🔥

Stay tuned for further analyses and experiments! ✨

A huge thanks to my brilliant collaborators for all their efforts and invaluable guidance: @eijkelboomfloor.bsky.social, @alisometry.bsky.social and @erikjbekkers.bsky.social 🔥

🧵5/5

We test our method on a checkerboard distribution on the sphere 🏁🌐, yielding promising results when compared to Euclidean and non-variational methods.

🧵4/5

By directly encoding the manifold, we introduce Riemannian-Gaussian-VFM, offering a fresh perspective on geometry-informed generative modeling. We also discuss how our approach relates to the previously mentioned methods. 🌍🌎🌏

🧵3/5

Inspired by Riemannian Flow Matching, we extend the Variational Flow Matching (VFM) objective to build a geometry-aware flow model. 🌊

Interestingly, VFM is particularly well-suited for explicitly parametrizing the support manifold of the distribution! 💅💪

🧵2/5

👮‍♀️🚨🚔

of course !!

we also celebrate it in some areas of Italy - like where I come from - where "Santa Lucia" is even more popular than Santa Claus for children!!

haters will say photoshop!!

So cooooool!!

Come and check out our poster on the UniReps workshop at @neuripsconf.bsky.social !

Thanks again to the amazing Giovanni Marchetti, Martina Scolamiero, and Danica Kragic ❤️

See more at: arxiv.org/abs/2409.10967

of course!:)