Collidify: Interactive Collider Decomposition for Convex Meshes and Primitives

Published August 16, 2026

Bitwise (pseudonym), independent researcher and developer of Collidify

Download the full paper ↓View on Zenodo ↗

Abstract

Automatic collider generation must balance decomposition quality, collider count, and runtime across convex meshes and geometric primitives. Existing decomposition pipelines typically target one representation, while downstream tools often convert convex decompositions to primitive colliders afterward. Collidify unifies convex and primitive decomposition in a greedy, GPU-batched framework. In paired median comparisons at fixed collider budgets, it approaches CoACD's quality with roughly 100× lower end-to-end runtime on the tested workstation. Each cut is selected by evaluating representation-specific prospective children: convex-mesh mode uses an immediate-progress reward, while primitive modes rank anticipated fit to the requested collider family. Approximate batched scoring and progressive candidate filtering enable broad candidate evaluation without look-ahead tree search.

On an external 461-model benchmark, Collidify approached CoACD's fixed-budget quality. At maximum budgets of 8, 16, and 32 colliders, its paired median relative neutral-concavity gaps against CoACD were +2.14%, +1.63%, and +0.84%; median speedups were 110×, 127×, and 111×. At the same budgets, VisACD's paired median relative neutral-concavity gaps against Collidify were +4.56%, +15.62%, and +29.85%, while V-HACD's were +15.49%, +39.56%, and +52.38%.

Controlled studies found that transplanting Collidify's immediate-progress reward into a CoACD-derived greedy shell reduced neutral concavity by 17.5% to 18.1% at fixed budgets of 16 and 32 hulls. For decompositions requested at 16 parts, GPU-scored decompositions had a paired median relative union-volume change of +0.047% against materialized exact CPU scoring and ran 30.3× faster. Discrete orientation search improved box and capsule quality-complexity frontiers relative to principal-component-analysis-only fitting, while representation-conditioned sphere partitioning reduced final sphere-union volume by 13% to 16% relative to a generic partition baseline.

Introduction

Real-time physics systems use simplified collision proxies to reduce simulation cost. Convex meshes provide flexible geometric approximation, while boxes, capsules, and spheres impose stronger shape constraints that can reduce processing cost and simplify manual editing. Automatic collider generation must therefore balance decomposition quality, collider count, runtime, and the requested output representation.

Top-down decomposition methods select an active region, evaluate candidate cuts, and commit to a partition until a quality threshold or part-count limit is reached. For primitive outputs, partitioning and orientation fitting are coupled: a cut suited to convex hulls may yield poor boxes or capsules. Collidify uses parallel evaluation of immediate, representation-specific decisions, then a CPU-controlled greedy loop to commit each selected cut.

Contributions

• An immediate-progress convex-mesh objective that selects locally useful cuts without look-ahead tree search.
• A GPU-batched approximate scorer with progressive candidate filtering for broad, low-latency candidate evaluation.
• Representation-conditioned cut selection that evaluates prospective child fit as convex meshes, boxes, capsules, or spheres before partitioning.

Method

Greedy, representation-specific selection

For every candidate plane, Collidify estimates both prospective children in the requested collider family before materializing a cut. Convex-mesh mode evaluates approximate support hulls. Primitive modes fit boxes, capsules, or spheres and evaluate their occupancy against the original source mesh. The selected plane is materialized; lower-ranked candidates remain fallbacks when a higher-ranked slice fails.

Convex-mesh mode combines immediate convex-envelope-volume reduction with a source-volume-weighted residual-work term. The result is normalized against the active parent hull. This preserves a greedy local objective: it does not estimate future cuts, allocate a fixed collider budget globally, or guarantee a globally optimal decomposition.

Candidate generation and GPU scoring

Candidate planes combine concave crease-edge samples with world-axis-aligned planes. The GPU evaluates prospective children without constructing a separate mesh for every plane. Broad candidate sets first receive an inexpensive score; only the best candidates advance to richer support sets and larger deterministic sample prefixes. This progressive cascade trades a small, configurable fidelity risk for substantially lower latency.

Primitive fitting

Box and capsule modes search discrete orientations for each prospective child, allowing the two children to retain independent orientations. Sphere mode uses deterministic support-point bounds to select a center and enclosing radius. The temporary fits participate in cut selection; after a cut is committed, final refitting and rollback preserve enclosure and the requested quality threshold.

Evaluation

External benchmark

The external comparison evaluates CoACD, Collidify mesh mode, VisACD, and V-HACD through a common orchestration and scoring layer. The 461-model cohort contains 200 Objaverse models, 200 ShapeNet models, and 61 models from the V-HACD evaluation dataset. Inputs were centered and scaled to a unit bounding-box diagonal; feasible outputs required coverage of at least 0.995. The external scorer reports neutral concavity, coverage, intersection over union, relative union-volume error, redundant overlap, Hausdorff distance, and Chamfer distance.

CoACD retains the stronger aggregate quality-part-count frontier when quality takes priority over latency. Collidify targets competitive median fixed-budget quality at latency suited to iterative authoring workflows.

Controlled studies

Controlled Unity studies isolate the effects of the mesh reward, approximate GPU scoring, discrete orientation search, representation-conditioned primitive partitioning, progressive scoring, and candidate-plane families. These mechanism studies use 40 models and model-level bootstrap confidence intervals. Primitive results are internal baseline comparisons, not claims of state-of-the-art primitive-decomposition quality.

Paper and software availability

The complete methodology, experimental design, limitations, supplementary figures, and references are available in the research paper.

Download the full paper ↓View on Zenodo ↗