GroomLab

Performance

Benchmarks of GroomLab against XGen Interactive Grooming at 100k and 500k strands, with the measurement method, caveats and how to rerun them.

GroomLab and Maya's XGen Interactive Grooming (Interactive Groom Splines) were timed on the same scalp and machine, at 100,000 and 500,000 strands. The benchmark harness is in the GroomLab source (tools/bench/xgen_compare), so you can rerun it on your own hardware.

  • Machine: AMD Threadripper PRO 5945WX (12 cores), 128 GB RAM, RTX 3090; Maya 2022, Windows 11.
  • Setup: The same 17k-triangle head-sized scalp, strands 15 cm long with 16 points each, clumping and noise on both (GroomLab adds a Curl), every strand shown in the viewport.
  • Runs: Medians of 3 to 5 runs, each measurement in a fresh Maya; the orbit turns about the framed head, as the tumble tool does (the GPU preview rows: 3 runs, 2026-10-02). XGen measured 2026-10-21, GroomLab 2026-10-23 (build 0.27.0).

Summary

  • Creation and file size. A new groom is evaluated in under a second, against about twelve seconds for XGen. A saved scene stores GroomLab's guides and graph, not its strands: a few MB at any strand count.
  • Animation and exports. With Follow Scalp the groom is evaluated once on the rest pose and transformed by the scalp, so per-frame evaluation of an animated scalp, Alembic export and motion-blurred Arnold export take a fraction of XGen's time.
  • Interactive edits. A parameter change redraws faster than XGen's at both strand counts, and the evaluation alone takes about half of XGen's time. While a slider or brush is dragged, the preview is evaluated on the GPU: about 12 ms to the redrawn frame at 500,000 strands, against 132 ms for an XGen tweak; on release, the result is computed on the CPU.

500,000 strands

MeasureXGenGroomLabFollow ScalpNo CurlGPU previewGroomLab vs XGen
New groom to first drawn frame13.3 s0.57 s–––23× faster
Saved scene (.mb)114 MB2.24 MB–––51× smaller
Animated scalp, evaluation per frame419 ms162 ms23.5 ms––18× faster (Follow Scalp)
Alembic export, 24 frames17.7 s6.89 s3.45 s––5.1× faster (Follow Scalp)
Arnold .ass export, motion blur8.24 s1.82 s1.01 s––8.2× faster (Follow Scalp)
Memory after creation1,127 MB661 MB–––1.7× less
GPU memory, viewport960 MB447 MB–––2.1× less
Save the scene (.mb)2.07 s0.02 s–––103× faster
Open the scene and evaluate1 s0.29 s–––3.4× faster
Clump tweak, evaluated252 ms103 ms–60.4 ms–2.4× faster
Noise tweak, evaluated162 ms75.6 ms–47.2 ms–2.1× faster
Noise tweak to redrawn frame132 ms91.7 ms–61.2 ms–1.4× faster
Clump tweak to redrawn frame402 ms116 ms–74.4 ms–3.5× faster
Dragging the clump slider, to redrawn frame402 ms–––13.3 ms30× faster (GPU preview)
Dragging the noise slider, to redrawn frame132 ms–––12 ms11× faster (GPU preview)
Animated scalp playback (higher is better)1.46 fps5.61 fps94.6 fps––65× faster (Follow Scalp)
Orbiting the viewport (higher is better)116 fps184 fps–––1.6× faster
Brush dab while dragging, to redrawn frame–57.8 ms–41.9 ms5.14 ms–

100,000 strands

MeasureXGenGroomLabFollow ScalpNo CurlGPU previewGroomLab vs XGen
New groom to first drawn frame14.6 s0.28 s–––52× faster
Saved scene (.mb)24.4 MB2.24 MB–––11× smaller
Animated scalp, evaluation per frame129 ms40.4 ms7.81 ms––17× faster (Follow Scalp)
Alembic export, 24 frames4.75 s1.31 s0.58 s––8.2× faster (Follow Scalp)
Arnold .ass export, motion blur2.36 s0.48 s0.29 s––8.1× faster (Follow Scalp)
Memory after creation286 MB192 MB–––1.5× less
GPU memory, viewport191 MB146 MB–––1.3× less
Save the scene (.mb)0.64 s0.02 s–––32× faster
Open the scene and evaluate0.49 s0.12 s–––4.1× faster
Clump tweak, evaluated52.4 ms23 ms–14.5 ms–2.3× faster
Noise tweak, evaluated33.2 ms17 ms–11.2 ms–2× faster
Noise tweak to redrawn frame36.1 ms25.6 ms–21.8 ms–1.4× faster
Clump tweak to redrawn frame97.6 ms33.4 ms–28.1 ms–2.9× faster
Dragging the clump slider, to redrawn frame97.6 ms–––4.08 ms24× faster (GPU preview)
Dragging the noise slider, to redrawn frame36.1 ms–––3.5 ms10× faster (GPU preview)
Animated scalp playback (higher is better)4.78 fps20.1 fps92.1 fps––19× faster (Follow Scalp)
Orbiting the viewport (higher is better)491 fps525 fps–––1.1× faster
Brush dab while dragging, to redrawn frame–19.2 ms–15.9 ms3.44 ms–

Follow Scalp is GroomLab with Follow Scalp on, for the measurements with an animated scalp. No Curl is the benchmark groom without its Curl node, so it runs the same operators as XGen here (clump and noise), for the closest comparison. GPU preview is GroomLab while a slider or brush is being dragged: the preview is evaluated on the GPU, and the CPU result on release differs from it in under 0.1% of pixels. Measured with Viewport 2.0 on OpenGL, as for XGen.

How it was measured

  • Every measurement runs in a fresh Maya: mayapy for batch work (creation, tweaks, exports, per-frame evaluation without drawing), the Maya GUI for the viewport (the window pinned to 1600 x 1000, the perspective view alone, smooth shaded, default lighting).
  • Creation is from nothing to a fully evaluated groom (XGen's xgmCreateSplineDescription with its clump and noise modifiers; GroomLab's create_groom, guides from curves and its graph). In the GUI, to the first drawn frame.
  • Tweaks change the clump strength or noise amplitude, then force an evaluation (batch) or a redraw (GUI).
  • Animated scalp: a bend deformer keyed over 100 frames; per-frame evaluation, playback and exports over its frames.
  • Arnold: .ass export time, static and with 2-key motion blur.
  • Memory: the process's working set after creation; GPU memory is the whole device's change (nvidia-smi).

Caveats

  • The two systems do different work. XGen evaluates each spline through its modifiers; GroomLab scatters strands, interpolates them from about 500 guides and runs its graph. Clump and noise algorithms and units differ, and GroomLab's clumps are tighter at the same strength.
  • XGen 2022 has no per-strand curl, so the matched comparison is No Curl.
  • XGen's brushes take only mouse events (they cannot be scripted), so brush timings are GroomLab's alone.
  • XGen uses OpenCL in the GUI; GroomLab evaluates on the CPU, multithreaded across all cores (except the GPU preview during drags).

Run it yourself

With a GroomLab build and Maya 2022 (and MtoA for the render part):

"C:/Program Files/Autodesk/Maya2022/bin/mayapy.exe" tools/bench/xgen_compare/run_bench.py --runs 5

--counts, --parts batch,render,gui and --systems pick what runs; the results go to tools/bench/xgen_compare/results/ as JSON and a Markdown table. A full run takes two to three hours. The harness README documents every option.

On this page