SculptForge / Benchmarks / How fast? · What we're building

How fast is SculptForge 0.11.2?

Measured with SculptForge's own benchmark on the same machines, the slowest one first. Every number on this page comes with the steps to repeat it. If your computer gives a different answer, send it to us; we list results that disagree with ours too.

The measured numbers are being re-checked and come with the next update. We found that our test scans had no sculpted detail, which makes some timings look better than real work, so we're measuring again on scans with real surface detail before we publish anything. What 0.11.2 changes, in words: big brushes on dense scans finish sooner after you lift the pen, they follow the pen more closely on faster computers, and orbiting a very dense scan right after a stroke is smoother.

The machines, slowest first

A speed claim counts only if it holds on the first tier that has a number. Today no real integrated-graphics laptop is in the set; until a tester reports one, the low-end budget is simulated (below) and every result says which tier it is.

Low-end budget, simulated

An Apple M1 Pro laptop running Linux, limited to a low-end machine's budgets and OpenGL 3.3 (SF_HW_SIMULATE=floor), on its fast cores, or on its two efficiency cores for the harshest CPU case.

Laptop

The same Apple M1 Pro laptop (32 GB shared memory), Linux, no limits.

Desktop

Windows, AMD Ryzen 5 5600X, NVIDIA RTX 3060, 32 GB.

Integrated graphics laptop

No machine yet. Have one? Run the steps below and send the file.

What's still slow

Big brushes (120 and 240 px) on very dense meshes still trail the pen on every machine we test, the desktop included: the stroke lands, but late.
On the biggest brushes (240 px) the stroke now follows the pen more closely while you draw; the cost is a slightly longer finish after you lift the pen on fast GPUs.
On the low-end budget, orbiting a 112 million triangle scan right after a stroke is still below a smooth 30 frames a second: the whole base mesh is drawn every frame. Making that cheaper is being researched.
Opening a scan with a micro level takes seconds, not instantly, on every machine (measured on the subdivided Lucy, with no sculpted detail; a file with sculpted micro detail can take longer).
Lucy at 112 million triangles needs more memory than an 8 GB integrated-graphics laptop has; on those, the largest Lucy that fits is the 28 million triangle version.

Repeat it on your computer

Install 0.11.2
From the download page (free, no account). For the comparison, also unpack 0.11.0 (the portable zip or AppImage runs beside the installed version). Close builds, games and renders.
Get the test scene
The built-in heads need nothing. For Lucy, download lucy.tar.gz (about 320 MB) from the Stanford 3D Scanning Repository and build the test scenes with the command-line tool that comes with SculptForge (below). We don't redistribute the scan.
Run the same strokes
Each run opens the scene, draws scripted Clay strokes at 25, 120 and 240 px through the real pen code, then orbits, and writes a JSON result and a video.

The exact commands

# the built-in 15 million triangle head (identical on every computer)
sculptforge --benchmark standard-15m --benchmark-out head-0.11.2.json

# Lucy: build the test scenes once from Stanford's lucy.ply (28 million triangles; 112 million with a micro level)
tar xzf lucy.tar.gz
sculptforge-cli -c "open path=lucy.ply; save_project path=lucy-28m.sfproj"
sculptforge-cli -c "open path=lucy.ply; micro_divide level=1; save_project path=lucy-112m.sfproj"

# Lucy: the same strokes and orbit on the scene
sculptforge --benchmark-file lucy-112m.sfproj --benchmark-out lucy-0.11.2.json

# the head with 0.11.0 (path to the unpacked 0.11.0)
./sculptforge-0.11.0/sculptforge --benchmark standard-15m --benchmark-out head-0.11.0.json

# optional: the simulated low-end budget we report first
SF_HW_SIMULATE=floor SF_GOVERNOR=0 sculptforge --benchmark-file lucy-28m.sfproj --benchmark-out lucy-floor.json

On macOS the command-line tool is SculptForge.app/Contents/MacOS/sculptforge-cli; on Windows it is sculptforge-cli.exe in the install folder. On Windows, set the variables first in a Command Prompt (set SF_HW_SIMULATE=floor) and run sculptforge.exe from the install folder. Run each build at least twice, alternating, and take the median. 0.11.0 has no --benchmark-file (it came in 0.11.1), so our 0.11.0 Lucy rows use 0.11.0 with only the benchmark's own commits added: the patch that lets 0.11.0 load a project file for the benchmark (no app code changed).

Reading the result

Stroke lag (lag_ms): from the pen sample to the frame on screen that shows its dab. p50 is the typical dab, p99 the slow ones you feel.
Release (release_done_ms): from pen up to the whole stroke in the mesh and shaded.
Orbit: frames per second while the view turns after the strokes, while the shading catches up.
Memory: peak RAM plus GPU buffers.

Full method, every tier and the raw numbers: the performance scoreboard. Format of the JSON file: docs/benchmark.md. Send yours: submit a result.

Comparing with other apps

This page lists SculptForge's numbers only. We don't run, publish or test other sculpting applications. Lucy is a public scan, so you can open the same model elsewhere and judge for yourself.