Collidify: Interactive Collider Decomposition for Convex Meshes and Primitives

Automatic collider generation must balance fit quality, collider count, and runtime across convex meshes and geometric primitives. Many existing decomposition methods select partitions without accounting for the requested output representation and fit primitives only after decomposition. We present Collidify, a greedy, GPU-batched method that evaluates representation-specific prospective children before committing to each cut. Convex-mesh mode ranks cuts with an immediate-progress reward, while primitive modes rank them by prospective fit to the requested collider family. Approximate batched scoring and progressive candidate filtering make broad candidate evaluation practical 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 gaps in neutral concavity relative to CoACD were +2.14% [0.70%, 4.46%], +1.63% [0.33%, 4.50%], and +0.84% [−1.60%, 3.70%], where brackets give pointwise 95% model-bootstrap confidence intervals. Median speedups were 110× [95.5×, 124.5×], 127× [121.7×, 136.3×], and 111× [106.7×, 117.8×]. Across the same budgets, VisACD and V-HACD produced paired median neutral concavities 4.56%–29.85% and 15.49%–52.38% higher than Collidify's, respectively. Controlled studies isolated the effects of Collidify's main design choices. Transplanting the immediate-progress reward into a CoACD-derived greedy shell reduced neutral concavity by 17.5%–18.1% at fixed budgets of 16 and 32 hulls. For decompositions requested at 16 parts, approximate GPU evaluation changed the median union volume by +0.047% compared with materialized exact CPU evaluation and was 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%–16% relative to the generic partition baseline. Collidify therefore targets competitive median fixed-budget quality at latency suited to iterative workflows; look-ahead search remains advantageous when aggregate quality–part-count performance takes priority over latency.