Testing 4 Free Mesh Simplification Tools
Lukáš Gallo shared with us the research he did in collaboration with Bohemia Interactive, explaining how he ran some performance tests on four open-source simplification tools, showing the different results of each one and the situations in which they can be used.
Introduction
I benchmarked four free, open-source mesh simplification tools (meshoptimizer, CGAL, Open3D, and fast-simplification) on production game assets from Bohemia Interactive's Arma Reforger to find out which one indie developers should actually use. I ran 288 benchmark operations across eight assets, then had 150 people judge visual quality in nearly 2,800 pairwise comparisons. Below is the recommendation table I wish had existed when I started, and the stories the numbers hide.
The tool with the lowest measured geometric error ranked third in what people actually preferred looking at. The tool people preferred most was also the fastest. And the tool that was second fastest collapsed a church tower into jagged spikes when asked to simplify it by 80%.
Why This Matters
If you have the budget, you license Simplygon and move on. But most of us do not work at studios where that is an option. If you are an indie developer, a student, or a small team, you are choosing between free tools with no empirical guidance on which one actually produces good results on real game assets.
The Four Tools
All four implement variations of the quadric error metric, the standard algorithm behind most mesh simplification. meshoptimizer, Open3D, and fast-simplification all install via pip in under three minutes. CGAL requires C++ compilation through vcpkg with 109 dependencies and roughly an hour of build time.
When to Use What
The top three tools are separated by only 3.4 percentage points in perceptual preference. The margins are tight. If you cannot install CGAL, meshoptimizer covers most of the same ground.
What the Table Does Not Show You
Numbers flatten out the stories. Here is what the data actually looked like.
fast-simplification uses an aggressive error threshold that escalates with each iteration. On compact, watertight geometry, it works fine. On the Church, with its tower spires and window mullions, the algorithm hit a wall at 79.5% reduction instead of the requested 90%, and what it did produce was structurally destroyed.
It also filled in the Jeep's windshield void and removed paddle fins from the Watermill. The developer's own documentation says the method "works well for watertight volumes without thin sections." That is a narrow niche, and most production pipelines do not fit in it.
Open3D tells a subtler story. It places vertices closer to the original surface than any other tool, by a factor of three. But it averages vertex normals when it collapses edges, which means hard-surface assets come out looking pillowed: smooth where the artist intended sharp edges. On organic geometry like the Bunker sandbags or the Mi-8 helicopter fuselage, that same smoothing actually helps, and Open3D led perceptually on those assets.
The problem is that its geometric error metrics look perfect while the visual result does not. That brings up the broader finding: geometric accuracy metrics like Hausdorff distance explained only about 10-15% of what drove players' visual preferences. Shading preservation, silhouette integrity, and feature retention mattered far more, and no standard metric captures those. If you are evaluating simplification tools by which one reports the lowest error number, you are probably optimizing for the wrong thing.
Meshoptimizer does not win dramatically on any single asset, but it never loses badly either. It classifies vertices by type before collapsing, automatically detects attribute discontinuities from the index buffer, and preserves shading boundaries that the other tools either ignore or average away. That consistency across asset types is what pushed it to the top in perceptual preference (57.6%), and at 24 ms average, it was also the fastest by a wide margin.
CGAL is the strongest at moderate reduction on compact hard-surface objects, leading perceptually at 50% reduction with a 63.5% win rate. But it degrades at aggressive levels, and installing it is a genuine barrier for anyone without C++ experience.
Try It Yourself
The complete benchmark pipeline, perceptual study platform, and analysis scripts are published on [GitHub]. You can run the same evaluation on your own assets and methods without rebuilding the infrastructure. The full thesis is included in the repository.
