Faithful Contouring: Near-Lossless 3D Voxel Representation Free from Iso-surface

Faithful Contouring addresses limitations in traditional 3D voxel representations (SDF + Marching Cubes), which often introduce artifacts like surface thickening and loss of detail. This project introduces a near-lossless, compact, and flexible voxel representation using "Faithful Contour Tokens" (FCTs) that directly operates on raw meshes. It benefits researchers and developers requiring high-fidelity, scalable 3D geometry processing, particularly for complex or non-manifold shapes.

How It Works

The core innovation avoids Signed Distance Fields (SDF) and Marching Cubes by identifying surface-intersecting voxels and solving for local FCTs. An encoder leverages a hierarchical octree traversal, SAT polygon clipping, Quadric Error Functions (QEF) for anchor points/normals, and segment-triangle intersection for edge flux signs. The decoder reconstructs the mesh by identifying edges with non-zero flux and performing adaptive triangulation based on normal consistency. This method preserves sharp edges and internal structures, even for open or non-manifold meshes, offering superior fidelity and scalability.

Quick Start & Requirements

The recommended installation uses Pixi (pixi run demo), which automatically handles dependencies including Python 3.10+, PyTorch 2.5+ (CUDA-compatible), torch_scatter, Atom3d, trimesh, scipy, and einops. A manual setup guide is also provided. The demo can be run with python demo.py, supporting custom meshes via the --mesh_path argument and adjustable resolution (-r). Links to the Pixi installer and the GitHub repository are available.

Highlighted Details

• Achieves near-lossless 3D reconstruction, preserving sharp edges and internal structures.
• Scalable to high resolutions (2048+) with efficient GPU kernels, demonstrated by benchmarks on NVIDIA H100.
• Features a compact voxel token representation (18 dimensions) supporting filtering, texturing, manipulation, and assembly.
• Accepted as an Oral presentation at CVPR 2026.
• Offers a pure Python + Atom3d implementation, eliminating the need for C++ compilation.

Maintenance & Community

The project is led by authors affiliated with multiple research institutions. While specific community channels (like Discord/Slack) are not detailed, concurrent work (TRELLIS 2) is noted, indicating active development in the broader research area.

Licensing & Compatibility

Licensed under the Apache License 2.0, this project is compatible with commercial use and integration into closed-source applications.

Limitations & Caveats

The README does not explicitly detail limitations such as alpha status or known bugs. Performance benchmarks indicate increasing decode times at higher resolutions (e.g., 2.51s for 2048 resolution), which may be a factor for real-time applications.