Mesh Silksong: Auto-Regressive Mesh Generation as Weaving Silk
This addresses a bottleneck in 3D mesh generation for applications requiring geometric integrity, though it appears incremental relative to existing tokenization methods.
The paper tackles the problem of inefficient mesh tokenization in auto-regressive mesh generation by introducing Mesh Silksong, which reduces token sequence redundancy by 50% and achieves a state-of-the-art compression rate of approximately 22% while producing meshes with superior geometric properties.
We introduce Mesh Silksong, a compact and efficient mesh representation tailored to generate the polygon mesh in an auto-regressive manner akin to silk weaving. Existing mesh tokenization methods always produce token sequences with repeated vertex tokens, wasting the network capability. Therefore, our approach tokenizes mesh vertices by accessing each mesh vertice only once, reduces the token sequence's redundancy by 50\%, and achieves a state-of-the-art compression rate of approximately 22\%. Furthermore, Mesh Silksong produces polygon meshes with superior geometric properties, including manifold topology, watertight detection, and consistent face normals, which are critical for practical applications. Experimental results demonstrate the effectiveness of our approach, showcasing not only intricate mesh generation but also significantly improved geometric integrity.