High-performance CUDA mesh processing library with BVH acceleration for voxelization, distance fields, mesh extraction, and intersection queries.

• [2026-01] atom3d.mesh_extractor.SparseDiffDMC — Differentiable mesh extraction from sparse voxel input
• [2026-01] atom3d.apps.VisibilityQuery — Ray-based visibility testing
• [2026-01] atom3d.apps.sparse_flood_fill — High-performance CUDA flood fill from multi-resolution voxelization

Atom3D provides high-performance primitives for modern 3D deep learning:

• Core Geometry: MeshBVH with SAT intersection, clipping, and UDF/SDF queries.
• Mesh Extraction: SparseDiffDMC for differentiable, sparse-grid isosurface extraction.
• Spatial Indexing: OctreeIndexer and CubeGrid for efficient sparse data management.
• Applications: VisibilityQuery and CUDA-optimized sparse_flood_fill.

👉 See Detailed Features Documentation

pip install -e . --no-build-isolation

or

pip install git+https://github.com/Luo-Yihao/atom3d.git --no-build-isolation

Requirements: Python ≥ 3.8, PyTorch ≥ 2.0, CUDA ≥ 11.0

from atom3d import MeshBVH
from atom3d.grid import OctreeIndexer
from atom3d.mesh_extractor import SparseDiffDMC

# BVH-accelerated mesh queries
bvh = MeshBVH(vertices, faces, device='cuda')
result = bvh.udf(points, return_closest=True)

# Octree-based voxelization
octree = OctreeIndexer(max_level=10, device='cuda')
candidates = octree.octree_traverse(bvh, min_level=4)

# Differentiable mesh extraction
dmc = SparseDiffDMC(device='cuda')
mesh_verts, mesh_faces = dmc(voxel_coords, sdf, cube_idx, resolution)

• cubvh — BVH implementation reference
• FlexiCubes — Differentiable mesh extraction
• diso — Differentiable isosurface extraction

This project is licensed under the MIT License.