Proximity-Preserving Neural Subdivision
Hassan Ugail
cs.GR
2026-08-11
PNS augments Loop subdivision with a bounded, curvature-gated neural correction: 21% better near-feature fitting than Loop while provably staying inside its proximity envelope, where an unconstrained neural baseline flips faces and diverges within four refinement levels.
Subdivision is a workhorse of geometry processing: apply one fixed local rule repeatedly, and a coarse mesh converges to a smooth limit surface. Its entire value sits in structural properties: local, repeatable, analytically tractable. But a fixed stencil underfits localized features: sharp ridges and soft edges, where curvature concentrates, get smoothed away. Neural mesh refinement adapts to such features, but it has to work as an operator ; its own output gets fed back in for further refinement. Unconstrained vertex prediction collapses right there: however good the one-step fit, repeated application produces high-frequency artefacts and flipped faces, at which point it is no longer a subdivision operator. Hassan Ugail's position in this paper: learning can enter subdivision, but locked inside the cage of subdivision structure.
PNS never predicts vertices. It adds a correction on top of the Loop edge vertex: new vertex = Loop edge vertex + h²·γ·F·η(φ). Four components, each with one job.
The design philosophy: every guarantee is architectural and holds for any finite weights before training. Training only selects a useful member of the guaranteed operator class; an adversarial weight choice cannot break these properties. The network itself is tiny (three-layer MLP, 64 hidden units), trained with a weighted combination of signed-distance, normal-alignment, fairness, and a soft proximity regularizer. The regularizer is not the source of the bound, which lives in the architecture; it merely discourages spending the full envelope.
Thirty analytic Gaussian ridges with randomized parameters (20 train / 10 held out), against Loop and an unconstrained neural baseline, paired bootstrap 95% CIs on the test set:
| Metric (vs Loop) | PNS | Unconstrained neural |
| Near-feature SDF RMSE | +21.0% [13.7, 31.3] | +56.0% [35.5, 69.4] |
| Global SDF RMSE | +13.3% [7.4, 21.4] | +54.0% [33.6, 67.5] |
| Global normal error | +0.1% (n.s.) | −4.0% (worse) |
| Max proximity ratio after 4 levels | ≤1 | 192.7 |
| Normal jump after 4 levels | 0.130 | ≈π (face flips) |
The unconstrained baseline is indeed the stronger one-step fitter, which is the paper's point: free vertex prediction has more capacity, and the same capacity disqualifies it as a subdivision operator. After four refinement levels its proximity ratio hits 192.7, two orders of magnitude past the envelope, and its normal jump approaches π, indicating extensive face flips; PNS stays at or below the architectural cap of 1 throughout, with a normal jump of 0.130, slightly better than Loop's 0.139. Classical adaptive schemes also fail the envelope test: Modified Butterfly lands at 61.6, the 1-ring quadric fit at 458.7. Budget allocation checks out too: per-edge proximity ratios lie on the identity line against the gate, with corrections concentrating on ridge-parallel edges while perpendicular edges get essentially none. Sweeping C from 0.1 to 1.0 moves near-feature improvement from roughly 5% to 22%, diminishing beyond 0.5, and the proximity ratio never breaks the cap at any C. On a smooth saddle dataset PNS gives no improvement over Loop, exactly the designed behavior where the gate stays closed. The code ships as a Blender add-on.
For geometry-processing practitioners this is a clean template for putting learning into a classical pipeline: not an end-to-end replacement, but a correction term with theoretical guarantees, leaving every structural property of the classical scheme intact. The C knob has a crisp semantics, and in production asset pipelines controllability beats peak accuracy. For AI practitioners at large, it demonstrates a pattern now recurring across fields: bounded, gated, equivariant corrections layered on a classical operator, with training only choosing within that subspace, so guarantees and performance stop being a trade-off. The gains themselves are incremental, and the 21% near-feature number rests on a synthetic benchmark tailored to ridges.
The authors are blunt: corrections scale as h², so under repeated refinement PNS becomes increasingly Loop-like ; it delivers finite-level improvement, not a new limit surface. Spectral inheritance is proven only at the planar reference; no global C¹ theorem at non-planar extraordinary vertices, where behavior is bounded but not quantified. The construction is tied to Loop and triangle connectivity; Catmull-Clark and quad meshes are natural extensions, unimplemented. Boundary edges route through Loop's boundary midpoint with no learned correction. Evaluation covers synthetic localized smooth features ; sharp discontinuities, noisy scans, and broad low-curvature regions are untested, and there the gate either saturates or closes, reverting the method to Loop. One more from reading the paper: the main experiments stop at four subdivision levels; the h² scaling gives a reason to extrapolate, but the extended-iteration study the authors themselves call for is missing.