trueform

trueform benchmarks

Performance, measured.

Every published trueform benchmark, in one place. Same inputs. Equivalent work. Every result validated. Each figure states the hardware, compiler settings, library versions, and timed boundary.

On the stated Apple M4 Max system (16 threads, trueform 0.10.10; mesh decimation, trueform 0.10.11), trueform records the fastest validated result of the libraries we benchmarked for mesh booleans, connected components, Delaunay/CDT, polygon arrangements, principal curvatures, isocontours, spatial tree builds, closest-point queries, mesh decimation, and mesh I/O. Each result is grouped under the module that owns the operation in trueform.

The benchmark harnesses, definitions, and published results are available in the trueform-benchmarks repository.

trueform module

spatial

Build spatial indexes and query geometry at scale.

Spatial tree build benchmark — C++

trueform builds a query-ready AABB tree over the 1.03M-face dragon in 7.87 ms, versus 34.4 ms for MeshLib and 177 ms for CGAL: 4.4× and 22× faster. The mesh is in memory before the clock; each library’s mesh conversion is not timed. trueform and MeshLib parallel; CGAL serial. Apple M4 Max, 16 threads, trueform 0.10.10.

Closest point

100K nearest-surface queries per mesh with the spatial tree prebuilt: build once, query many.

TutorialsClosest point on a mesh in C++Closest point on a mesh in PythonClosest point on a mesh in JavaScript

2.2× faster than MeshLib
Closest point benchmark — C++

100K closest-point queries per mesh on 160 Thingi10K meshes, tree prebuilt. At the ~1M-face group trueform answers the batch in 94.1 ms (geometric mean of 40 meshes): 2.2× faster than MeshLib and 2.3× faster than libigl. All three run parallel. Apple M4 Max, 16 threads, trueform 0.10.10.

trueform module

topology

Inspect and traverse mesh connectivity.

Connected components

Label every edge-connected component of a triangle mesh from the connectivity a held mesh keeps.

TutorialsConnected components in C++Connected components in PythonConnected components in JavaScript

up to 7.9× faster
Connected components benchmark — C++

On 160 Thingi10K meshes in four size groups, trueform labels every edge-connected face in 2.38 ms at the ~1M-face group (geometric mean of 40 meshes): 4.9× faster than libigl and 7.9× faster than CGAL. The face link, and each peer’s mesh and adjacency, are built before the clock. trueform parallel; CGAL and libigl serial. Apple M4 Max, 16 threads, trueform 0.10.10; every result validated.

1M points in 18.0 ms
Delaunay & CDT benchmark — C++

On 1M random points, trueform completes Delaunay triangulation in 18.0 ms: 1.5× faster than Delaunay32 and 25–46× faster than Triangle, CGAL, detria, Geogram and artem-ogre/CDT. On 1M points with closed-contour constraints (about 100K segments), the triangulation takes 50.1 ms: 9.4× faster than Triangle and 12–17× faster than CGAL, Geogram, detria and artem-ogre/CDT. Full convex hull for every library; region labels excluded; every row validated. trueform and Delaunay32 parallel, the others serial as shipped. Apple M4 Max, 16 threads, trueform 0.10.10.

trueform module

geometry

Measure and evaluate geometric properties.

Principal curvatures

Estimate both principal curvature values and directions at every vertex of an indexed surface.

TutorialsPrincipal curvatures in C++Principal curvatures in PythonPrincipal curvatures in JavaScript

85× faster than libigl
Principal curvatures benchmark — C++

On indexed spheres from 57K to 1.02M faces, trueform computes both principal curvature values and directions in 20.0 ms at the largest size versus 1.70 s for libigl: 85× faster. The normals and neighbourhoods a held mesh keeps are built before the clock; libigl rebuilds its adjacency inside every call. trueform parallel, libigl serial. Apple M4 Max, 16 threads, trueform 0.10.10.

Geodesic distances

Distance along the surface from one seed or 32, on 28 Thingi10K meshes from 10K to 3.2M faces, checked against exact MMP geodesics.

TutorialsGeodesic distances in C++Geodesic distances in PythonGeodesic distances in JavaScript

6.5× faster than MeshLib
Geodesic distances benchmark — C++

Over 28 Thingi10K meshes from 10K to 3.2M faces, trueform 0.10.8 computes geodesic distances from 32 seeds in a 6.0 ms geometric mean versus 39.4 ms for MeshLib — 6.5× faster, and 4.4× from one seed. Its median error against exact MMP geodesics is 0.57%; MeshLib's is 2.3%. The heat methods of libigl and CGAL run 47–99× slower below 1M faces, at 13–36% error. Mesh and topology are built inside the clock.

trueform module

remesh

Change mesh resolution while preserving shape.

Mesh decimation

Edge-collapse to 10% of the faces, manifoldness and boundary preserved, on 160 Thingi10K meshes.

18.5× faster than CGAL
Mesh decimation benchmark — C++

Decimation to 10% of the faces on 160 Thingi10K meshes. At the ~1M-face group trueform takes 177 ms (geometric mean of 40 meshes): 1.07× faster than MeshLib (189 ms) and 18.5× faster than CGAL (3,280 ms). trueform uses a quality floor of q ≥ 0.3 and MeshLib its equivalent maxTriangleAspectRatio of 2.13855; CGAL has no equivalent hard floor and runs serially. Apple M4 Max, 16 threads, trueform 0.10.11.

trueform module

arrangement

Split geometry at intersections and recover the resulting structure.

Polygon arrangements

Split a self-intersecting polygon soup along every intersection curve and return the complete subdivided mesh.

6.6–8.2× faster than CGAL
Polygon arrangements benchmark — C++

On a self-intersecting soup made from two copies of the dragon, 2.05M input faces, trueform splits every intersection and returns the subdivided mesh with its face labels in 202 ms: 8.2× faster than CGAL (parallel) and 43× faster than libigl. Its one-off prep, the tree and edge topology a held mesh keeps, takes 51 ms more; counting it, the ratios are 6.6× and 34×. CGAL cannot prebuild its search structure. Apple M4 Max, 16 threads, trueform 0.10.10.

trueform module

iso

Extract contours and bands from scalar fields on meshes.

Isocontours

Extract connected curve geometry where a per-vertex scalar field crosses one or more levels.

TutorialsIsocontours on a mesh in C++Isocontours on a mesh in PythonIsocontours on a mesh in JavaScriptMesh slicing in C++Mesh slicing in PythonMesh slicing in JavaScript

1M faces in 4.50 ms
Isocontours benchmark — C++

Across 16 scalar levels on the 1.03M-face dragon, trueform returns connected contour paths in 4.50 ms: 47× faster than VTK and 159× faster than libigl. VTK and libigl return unconnected line segments; trueform also connects them into paths. Each library’s mesh and the field are built before the clock. trueform and VTK parallel, libigl serial. Apple M4 Max, 16 threads, trueform 0.10.10.

Embedded isocontours

Split a surface along scalar-field contours and return one fully triangulated, band-labeled mesh.

TutorialsIsocontours on a mesh in C++Isocontours on a mesh in PythonIsocontours on a mesh in JavaScriptMesh slicing in C++Mesh slicing in PythonMesh slicing in JavaScript

1M faces in 8.17 ms
Embedded isocontours benchmark — C++

Embedding 16 scalar levels into the 1.03M-face dragon, trueform returns the band-labeled mesh in 8.17 ms versus 584 ms for VTK: 71× faster. With the connected contour curves also returned, 9.88 ms versus 582 ms: 59×. trueform and VTK parallel. Apple M4 Max, 16 threads, trueform 0.10.10.

trueform module

csg

Combine solid meshes with boolean operations.

Variadic mesh booleans

One arrangement over 4, 16, or 64 operands, compared with native N-ary and chained pairwise implementations.

TutorialsVariadic mesh booleans in C++

up to 648× faster
Variadic mesh booleans benchmark — C++

N-ary union over 4, 16, and 64 Thingi10K meshes. At N = 64, trueform completes the full arrangement in 103 ms — 14× faster than Solidean and Manifold, 138× MeshLib, and 648× Geogram.

Read the full benchmark

trueform module

io

Read and write common mesh formats.

Mesh I/O

OBJ and binary STL into indexed float32/int32 geometry, including exact vertex welding for STL.

TutorialsSTL and OBJ reading in C++STL and OBJ reading in PythonSTL and OBJ reading in JavaScript

1M OBJ in 5.76 ms
Mesh I/O benchmark — C++

At 1.03M triangles, trueform reads triangulated OBJ into indexed float32/int32 arrays in 5.76 ms: 1.4× faster than RapidObj, 9.9× tinyobjloader, 105× CGAL and 174× libigl. It reads and exactly welds binary STL in 35.8 ms, 5.9–9.2× faster than stl_reader, VTK, libigl and CGAL. Warm file in the OS cache; open, parse and output arrays timed; every row validated. Apple M4 Max, 16 threads, trueform 0.10.10.