Neural mesh refinement

• Introduces a neural mesh refinement (NMR) method that learns geometric structural priors from fine meshes to adaptively refine coarse meshes via subdivision, demonstrating robust generalization. • Key innovation: disentangling the network from non-structural information (scale, rotation, translation) using an intrinsic structure descriptor and a locally adaptive neural filter with graph attention. • The method outperforms existing subdivision methods in geometry quality on diverse complex 3D shapes, enhancing generalization to unseen shapes and arbitrary refinement levels. • Charbonnier loss is shown to alleviate over-smoothing compared to L2 loss, contributing to improved geometric learning.