Transferability of QHFlow improvements to other SE(3)-equivariant backbones

Determine whether the performance improvements of the high-order SE(3)-equivariant flow matching framework QHFlow observed when implemented on QHNet transfer when QHFlow is implemented on other SE(3)-equivariant neural network backbones such as WANet or SPHNet.

Background

QHFlow is introduced and evaluated in this paper using QHNet as the underlying SE(3)-equivariant backbone due to public code availability and manageable training cost. The authors note that, in principle, the framework should be compatible with other backbones such as WANet or SPHNet.

However, because those implementations are not open-sourced, the authors were unable to verify whether the observed gains from QHFlow are backbone-agnostic. Establishing transferability would clarify the generality of the proposed flow matching approach across different equivariant architectures.