Luo-Yihao/FaithC

What it solves

Faithful Contouring (FaithC) addresses the limitations of conventional voxel-based mesh representations that rely on Signed Distance Fields (SDF) and Marching Cubes. These traditional pipelines often suffer from surface thickening, jagged edges, and the loss of internal structures, especially when dealing with open or non-manifold meshes that require complex preprocessing.

How it works

Instead of using iso-surfaces, FaithC operates directly on the raw mesh to identify surface-intersecting voxels. It encodes these voxels into compact 18-dimensional Faithful Contour Tokens (FCTs), consisting of a surface representative point (anchor), a surface normal, and edge crossing signs (edge flux).

• Encoder: Uses hierarchical octree traversal and BVH-accelerated intersection to find intersecting voxels, then solves for optimal anchor points and normals using a Quadric Error Function (QEF).
• Decoder: Reconstructs the mesh by identifying edges with non-zero flux and forming quads from incident voxels, followed by adaptive triangulation.

Who it’s for

Researchers and developers working in 3D geometry processing, 3D generative AI (such as VAEs or Diffusion models), and anyone needing a high-fidelity, near-lossless voxel representation for 3D meshes.

Highlights

• Near-Lossless Reconstruction: Preserves sharp edges and internal structures for both manifold and non-manifold meshes.
• High Scalability: Efficient GPU kernels allow for resolutions up to 2048+.
• Compact Representation: Each voxel token is only 18 dimensions.
• Flexible Format: The token format supports filtering, texturing, and manipulation.