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, trueform records the fastest validated result in these comparisons for mesh booleans, Delaunay/CDT, closest-point queries, Python mesh decimation, and mesh I/O. It matches the leading C++ implementations in mesh decimation and ICP.
Benchmark 01
Mesh Booleans
Union, intersection and difference on the Thingi10K corpus at 100K–1M triangles per operand.
- C++
- Eight libraries on the Thingi10K corpus at 100K–1M triangles per operand. Fastest on every pair — 5.5× MeshLib, 31× CGAL, 105× Geogram at the geometric mean. Full benchmark
- Python
- Three pip-installable libraries, same corpus: an 18 ms median boolean at 200K–1.5M polygons — fastest on all 1000 pairs. Full benchmark
- TypeScript
- Client-side WebAssembly head-to-head: a 21.9 ms median boolean — 12.6× faster than Manifold WASM, closed and manifold on all 1000 pairs. Full benchmark
Benchmark 02
Variadic mesh booleans
One arrangement over 4, 16, or 64 operands, compared with native N-ary and chained pairwise implementations.
- 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. Full benchmark
- C++
- On 1M random points, trueform completes Delaunay triangulation in 26.0 ms — 1.4× faster than Delaunay32 and 17.1–28.6× faster than Triangle, CGAL, Geogram, artem-ogre/CDT, and detria. With 100K non-crossing constraints, trueform completes the CDT in 62.3 ms, 7.6–12.7× faster than the five participating libraries. Float/int32; arrays in → arrays out; region labels excluded; every row validated.
Benchmark 04
Closest point
100K nearest-surface queries per mesh with the spatial tree prebuilt — build once, query many.
- C++
- 100K nearest-surface queries per mesh on the Thingi10K corpus, tree prebuilt — a 73 ms median batch, 1.8× faster than MeshLib and 2.2× faster than libigl. Parallel broadcast queries, arrays in → arrays out.
- Python
- 100K nearest-surface queries per mesh on the Thingi10K corpus, tree prebuilt — a 76 ms median batch, ~2× faster than both MeshLib and libigl. Broadcast queries, arrays in → arrays out.
Benchmark 05
Mesh decimation
Quadric edge-collapse to 10% of the faces, manifoldness preserved — the hard part done right.
- C++
- Decimating Thingi10K meshes to 10% — a 66 ms median, on par with MeshLib and 10× faster than CGAL Garland–Heckbert. All three preserve manifoldness (the bulk of the work); arrays in → arrays out, including the half-edge build.
- Python
- Decimating Thingi10K meshes to 10% at a shared 0.3 quality floor — a 55 ms median, 1.7× faster than MeshLib and 2.3× faster than pyfqmr. Arrays in → arrays out, including the half-edge build.
Benchmark 06
ICP registration
Rigid point-cloud alignment — recover a known transform by registering a perturbed copy.
- C++
- Point-to-point registration with 10% sampling and an RMSE convergence gate — a 36 ms median, neck-and-neck with small_gicp (2024 SOTA GICP) and 1.4× faster than MeshLib. Arrays in → arrays out.
- Python
- Point-to-plane registration with 10% stochastic sampling per iteration and an RMSE convergence gate — a 52 ms median, on par with small_gicp and 1.3× faster than MeshLib, at the highest convergence rate of the three.
Benchmark 07
Mesh I/O
OBJ and binary STL into indexed float32/int32 geometry, including exact vertex welding for STL.
- C++
- At 1.03M triangles, trueform reads triangulated OBJ into indexed float32/int32 geometry in 6.54 ms — 2.0× faster than RapidObj, 10.2× tinyobjloader, 112.5× CGAL, and 145.5× libigl. It reads and exactly welds binary STL in 42.5 ms, 5.4–9.3× faster than stl_reader, VTK, libigl, and CGAL. Warm-file timings include opening, parsing, and final output construction; every row validated.