Reproduce “AnySurf: Any Surface Generation with Directed Edge” on OpenResearch

About this paper

Wenda Shi, Chenyuan Pan, Dengming Zhang, Yiren Song, Biao Zhang, Xingxing Zou

AnySurf unifies the generation of open, closed, and hybrid 3D surfaces by introducing a novel representation that explicitly encodes normal direction, resolving the persistent issue of inconsistent face orientation. The framework achieves state-of-the-art normal quality and geometric accuracy across various 3D asset categories, outperforming existing general and domain-specific methods.

Key insights

• The fundamental issue of inconsistent face orientation in previous general 3D generation methods for open surfaces stems from the core 3D representation (e.g., FDG) inherently discarding explicit normal information during the encoding process.
• Explicitly encoding directed edges within a dual-grid representation can effectively and robustly preserve normal orientation cues throughout the 3D generation pipeline, enabling accurate face orientation in reconstructed meshes.
• A lightweight, parallel Feature Pyramid Network (FPN) adapter combined with a progressive, co-training fine-tuning strategy allows for the effective injection of new capabilities like orientation awareness into large pre-trained generative models while maintaining their existing geometric and topological priors.

Summary by alphaXiv.