Mesh Simplification Algorithms Explained

Mesh simplification replaces a surface with fewer elements while trying to preserve properties that matter. “Preserve” might mean geometric distance, silhouette from allowed cameras, UV layout, material boundaries, skin deformation, volume, topology, or recognizable small parts. No algorithm can infer every production requirement from vertex positions alone.

Most real-time pipelines rely on edge-collapse methods, often with quadric error metrics, because they provide a practical balance of quality and speed. Other families remain useful for fast previews, terrain, scan cleanup, topology repair, or view-dependent systems. Choosing a method begins with the input assumptions and the failure that is unacceptable.

Simplification is an optimization problem#

An algorithm needs four definitions:

the operations it is allowed to perform;
a cost or error assigned to each candidate operation;
constraints that reject invalid or forbidden operations;
a stopping condition.

For edge collapse, the operation replaces two edge endpoints with one vertex and removes the triangles that become degenerate. The cost estimates damage. Constraints can lock boundaries or reject normal flips. The stop rule can target 20,000 triangles, 1 millimeter error, or the first candidate that exceeds a budget.

Changing any one definition changes the result. Two tools both labeled “50% decimate” can choose different edges, positions, boundary rules, and attribute handling. Their identical triangle totals do not imply identical quality.

Edge-collapse simplification#

Edge collapse is the dominant local method for triangle meshes. For every eligible edge, the simplifier proposes a replacement vertex. It estimates the cost, selects a low-cost candidate, validates topology and geometry, performs the collapse, and updates nearby candidates.

On an interior manifold edge, a normal collapse usually removes 2 triangles. Collapsing 40,000 such edges can remove about 80,000 triangles, with differences at boundaries and special topology.

The CGAL surface-mesh simplification manual describes a collection stage that assigns initial edge costs and a collapse stage that processes candidates in increasing cost order while recalculating affected neighbors. Some processed edges are discarded because they fail topological or geometric conditions.

Advantages include:

fine local control;
support for geometric and attribute-aware costs;
explicit boundary and feature constraints;
good progressive reduction behavior;
compatibility with priority-queue implementations;
production-quality results on many manifold surfaces.

Weaknesses include:

dependence on reliable local topology;
greedy decisions that are not globally optimal;
possible loss of thin disconnected features;
priority-queue and neighborhood-update complexity;
sensitivity to cost scaling and constraints;
no automatic understanding of semantic importance.

Edge collapse is a strong default for static props, scans, LOD generation, and controlled batch optimization when inputs are preflighted.

Quadric error metrics#

The classic Garland–Heckbert quadric error metric represents the squared distance of a candidate vertex to planes associated with the original local surface. Each vertex carries a compact symmetric matrix. When vertices combine, their quadrics add; evaluating a candidate position produces an estimated accumulated error.

QEM is efficient because it carries local surface evidence forward instead of comparing every candidate with every original triangle. It tends to remove edges on broad planar regions before edges that define corners or curvature. Boundary quadrics, attribute terms, and feature constraints improve production behavior.

The basic plane-distance metric does not directly know texture distortion, material identity, screen coverage, or whether a thin antenna is semantically important. Those require extended costs, locks, or downstream validation.

QEM deserves its own detailed treatment because “quadric simplification” encompasses many placement, topology, and attribute variants.

Vertex clustering#

Vertex clustering divides space into cells and maps vertices in the same cell to a representative. Triangles are rewritten to the representatives, then degenerates and duplicates are removed.

If a 1-meter object uses a 100 × 100 × 100 grid, each axis cell is about 10 millimeters. Geometry within a cell can merge regardless of whether an edge explicitly connects it, depending on the implementation. That makes clustering fast and capable of joining nearby components.

Advantages:

near-linear processing in practical implementations;
simple parallelization;
predictable spatial scale;
strong reduction for dense noisy input;
tolerance of some irregular connectivity;
useful coarse previews or far LODs.

Risks:

thin gaps can close;
unrelated nearby surfaces can weld;
cell alignment can bias the result;
topology may change dramatically;
fine features below the cell size disappear;
UV and material attributes need explicit handling.

Clustering is useful when speed matters more than close fidelity or when the desired error is naturally spatial. Test several grid origins or hierarchical approaches if grid alignment creates visible bias.

Vertex removal and retriangulation#

Some methods remove a vertex and retriangulate the hole left by its one-ring neighbors. The cost can estimate local distance, curvature, normal change, or triangle quality.

This operation can preserve existing neighboring vertices while changing connectivity. Retriangulation is more involved than collapsing one edge, especially for non-convex neighborhoods or boundaries. Poor choices create skinny triangles or change local shading.

It can be useful when a workflow wants to retain a subset of original sample positions. For general real-time LOD creation, edge collapse has broader tool support and simpler progressive interpretation.

Triangle removal is not enough#

Deleting isolated triangles without repairing the surface creates holes. A simplifier must change connectivity or replace surface regions. Some scan tools remove disconnected components by triangle count, but that is cleanup based on component classification, not surface approximation.

Randomly deleting 50% of faces from a closed 100,000-triangle mesh leaves a perforated object. The number is lower; the model is not a simplified surface. Production metrics must include validity and image behavior.

Remeshing and resampling#

Remeshing constructs a new tessellation rather than preserving most source connectivity. Isotropic remeshing tries to produce triangles with relatively uniform edge length. Adaptive remeshing uses finer triangles in high-curvature or feature regions and coarser triangles elsewhere.

Remeshing can improve triangle quality and remove pathological source density. It can also change vertex positions, UV topology, material boundaries, skin weights, morph targets, and baked tangent space. A scanned sculpt intended for a fresh bake is a better candidate than a finished skinned character with approved UVs.

Use remeshing when the pipeline can regenerate dependent data. If UVs and bakes must remain byte-stable, an attribute-preserving simplifier is safer.

Measure both approximation error and element quality. A uniform mesh with beautiful triangles can still erase a sharp logo edge; a geometrically accurate output can contain slivers that harm later work.

Voxelization and surface extraction#

Voxel workflows sample the model into a volume or signed-distance field, then extract a surface using marching cubes, dual contouring, or related methods. Resolution determines which gaps and features survive.

A 256³ dense grid contains 16,777,216 cells before sparse representations. Across a 2-meter object, one axis step is about 7.8 millimeters. Any feature near that scale needs close review.

Advantages include:

combining overlapping components;
filling certain holes;
producing a more regular closed surface;
ignoring chaotic original connectivity;
a natural spatial error scale.

Costs include:

high memory at fine resolution;
loss of UVs and original attributes;
closing intentional openings;
rounding sharp corners;
complete rebaking and material transfer;
altered topology and component identity.

Voxelization is a reconstruction method, not a drop-in decimate checkbox. It is powerful for scan cleanup, collision proxies, and assets whose original topology is unusable.

View-dependent and silhouette-aware methods#

Object-space error treats all viewing directions equally. A fixed product viewer may allow only a front arc; geometry on the hidden rear can be cheaper. View-dependent methods weight error by permitted cameras, projected size, or current view and can generate dynamic levels.

These methods can produce excellent results for constrained experiences, but they couple the asset to camera policy. The same output may fail in AR where the user walks around it.

A practical offline approach renders candidate collapses or completed meshes from a camera set and adds silhouette or image penalties. A runtime approach selects clusters or continuous LOD based on view. Both require more infrastructure than a standard asset LOD chain.

Document the camera distribution. “View-aware” without recorded views is not reproducible.

Attribute-aware simplification#

Geometry is only part of a rendered vertex. UVs, normals, tangents, colors, skin weights, and morphs can constrain or influence reduction.

An attribute-aware cost might penalize:

UV displacement in texture space;
normal or tangent deviation;
color change;
skin-weight difference;
crossing a material boundary;
destroying a seam or hard edge;
morph-target error.

Weights need normalization. If object coordinates span 10 meters and UVs span 0–1, raw numeric magnitudes are not comparable. A UV weight of 1 and position weight of 1 do not express equal perceptual importance automatically.

Locking every seam preserves appearance but can prevent meaningful reduction on a mesh with hundreds of tiny UV islands. Allowing every seam to move can cause texture swimming. Use constraints by asset class and validate the exported material.

Topology and geometric validity filters#

A low-cost operation can be invalid. Filters commonly reject collapses that:

create non-manifold edges or vertices;
flip triangle orientation beyond a threshold;
produce degenerate or near-zero-area triangles;
cross a protected boundary;
change genus or join components when forbidden;
move outside a geometric envelope;
invert a local volume;
violate locked vertices or attributes.

Stricter filters preserve structure but can stop before the target count. That is useful evidence: the target conflicts with constraints. A tool should report “stopped at 18,420 triangles” rather than silently breaking features to reach 10,000.

CGAL exposes policies and filters for cost, placement, constraints, normal change, and geometric envelopes. Other libraries expose different controls. Map tool settings to the underlying behavior before standardizing a preset.

Stop conditions#

Triangle ratio is simple but not quality-based. Better systems can stop on:

target edge, face, or index count;
maximum local collapse cost;
accumulated or normalized geometric error;
absolute object-space tolerance;
screen-space error at a defined projection;
time or memory budget;
inability to find another valid collapse.

Use both a resource target and an error ceiling. “Try for 20% triangles, but never exceed 2 millimeters absolute deviation or the approved silhouette threshold” prevents a small number from overriding quality.

Relative error is convenient across differently scaled assets only when normalization is defined. Absolute error is easier to connect to manufacturing or world scale. Screen error is closest to perception but depends on cameras and resolution.

Selecting a method by asset#

For a clean static prop with UVs and materials, start with attribute-aware edge collapse and protected boundaries. For a dense photogrammetry mesh that will be rebaked, compare QEM edge collapse with adaptive remeshing. For an unusable collection of intersecting scan fragments, consider voxel reconstruction before final simplification. For terrain, hierarchical or tile-aware error and crack control matter. For a fixed-view impostor, view-aware reduction may justify aggressive hidden-side changes.

For an animated character, use a method that understands skin weights or review every pose. Generic geometric decimation can preserve the bind silhouette while collapsing elbows and eyelids in motion.

Benchmark protocol#

Preserve and hash the source mesh.
Repair only defects required by the candidate algorithm.
Define topology, boundary, attribute, and camera constraints.
Generate a ratio ladder with each method.
Record processing time and peak memory.
Measure triangles, rendered vertices, components, and attributes.
Calculate geometric and screen-space error.
Render fixed views and moving highlights.
Test animation and runtime when applicable.
Compare the smallest passing candidate, not equal ratios alone.

Equal triangle count is useful for quality comparison. Equal visual error is useful for efficiency comparison. Report both.

Algorithm selection checklist#

Input manifold and boundary assumptions are known.
The allowed operation matches required topology preservation.
Error terms cover important geometry and attributes.
Feature, seam, material, and component constraints are explicit.
Stop conditions include a quality ceiling.
Failure to reach the count is reported, not hidden.
Ratios are derived from exported triangles and vertices.
Fixed cameras and lighting make results comparable.
Runtime structure and memory are measured after export.
Dependent UV, bake, skin, morph, and collision data are revalidated.
Source and candidate configuration are reproducible.
The selected method wins on the asset class, not one showcase model.

Sources and further reading#

Mesh simplification is not one algorithm or one percentage. It is a controlled search over permitted surface changes. Choose operations, error, constraints, and stopping rules that describe the asset's real job, then validate the result in the image and runtime that will ship.