Curvature-Difference Methods Overview

• Curvature-difference methods are a collection of techniques that adjust standard curvature computations to isolate reliable geometric signals from numerical artifacts.
• They address challenges in diffusion models, level-set interfaces, geometric flows, and VOF reconstructions by mitigating issues like overfitting, non-smoothness, and instability.
• These methods enhance accuracy and robustness in applications such as image generation, interface tracking, and adaptive mesh redistribution in geometric flow simulations.

Curvature-difference methods denote a heterogeneous set of techniques in which curvature, or a curvature-derived quantity, is estimated, regularized, or explicitly differenced in order to isolate a target geometric signal from confounding effects. In the recent literature summarized here, the term includes coordinate-wise Hessian subtraction for local memorization in diffusion models (Kim et al., 26 May 2026), kink-avoiding curvature evaluation in level-set methods [(Lervåg et al., 2014); (Ervik et al., 2014)], monotone and filtered finite-difference discretizations for affine-curvature flow (Oberman et al., 2016), semidiscrete curvature quotients for hyperbolic curvature flow (Deckelnick et al., 2022), adaptive monitor functions built from curvature and curvature variation for Willmore flow (Duan et al., 4 Jan 2026), and curvature estimation from volume fractions via height functions and paraboloid fitting (Han et al., 2023). A plausible unifying interpretation is that these methods do not treat raw curvature as intrinsically trustworthy; instead, they modify its computation so that the retained signal is less contaminated by overfitting, medial-axis kinks, transport error, or mesh distortion.

1. Conceptual scope

Across these works, the operative object is always a curvature-related observable, but the source of error differs sharply by domain. In diffusion models, the confounder is intrinsic data-driven rigidity, which can produce high curvature even without memorization. In interface tracking, the confounder is non-smoothness of the signed-distance field near points equidistant to multiple interfaces. In curvature-driven PDEs, the central issue is nonlinear instability or loss of monotonicity under naive central differencing. In adaptive geometric evolution, curvature and curvature variation are used as monitor variables for mesh redistribution rather than as a force alone. In VOF-based interface reconstruction, the main trade-off is between static accuracy, robustness to advected volume-fraction error, and computational cost.

Domain	Curvature quantity	Main corrective mechanism
Diffusion models	Diagonal Hessian of log-density	Subtract unconditional or less-trained baseline
Level-set interfaces	$\kappa=\nabla\!\cdot(\nabla\phi/|\nabla\phi|)$	Avoid stencils crossing kinks; reconstruct locally smooth fields
Geometric flows	Discrete curvature quotients or affine-curvature operators	Monotone wide stencils, semidiscrete ODEs, filtered schemes
Willmore flow	$\kappa$ and $\partial_s\kappa$	Monitor-based redistribution or tangential velocity
VOF interfaces	Mean curvature from heights or paraboloid fits	Estimator selection by accuracy-cost regime

This heterogeneity is important. A common misconception is that “curvature-difference” always means subtracting two curvature fields. That interpretation is exact for the diffusion-model construction, but the broader literature also uses the term for finite-difference curvature quotients, wide-stencil reconstructions, or local extraction procedures whose purpose is to prevent curvature evaluation from crossing a singular or non-smooth representation.

2. Coordinate-wise curvature differences in diffusion models

The most explicit curvature-difference framework appears in the study of local memorization in text-to-image diffusion models (Kim et al., 26 May 2026). At noise level $t$, with state $x_t\in\mathbb R^d$, the conditional and unconditional score networks are

$$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$

and the Hessian of the log-density is

$H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$

The coordinate-wise curvature is the diagonal entry

$$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$

which measures the “sharpness” along coordinate $i$.

The geometric motivation is a curvature-variance correspondence. On a Gaussian, directions of low variance $\sigma^2$ have high curvature $\kappa$0. The paper formalizes this through the curvature-variance identity

$\kappa$1

so that $\kappa$2 is, up to scale, the inverse conditional variance. Local memorization is therefore characterized as a coordinate-wise variance collapse. The complication is that low conditional variance can also arise from intrinsic structure in $\kappa$3, such as semantically rigid regions or simple textures. The method consequently defines Overfitting-Driven Memorization by subtracting a baseline curvature.

The conditional and baseline curvatures are

$\kappa$4

and the curvature-difference score is

$\kappa$5

The less-trained baseline is an earlier checkpoint, such as SD v1.1 for target SD v1.4. The unconditional baseline reuses the model’s unconditional U-Net used for classifier-free guidance.

A second construction gives a score-difference proxy with a Fisher-information interpretation. If

$\kappa$6

then, at late sampling steps $\kappa$7,

$\kappa$8

For a less-trained baseline, the proxy becomes

$\kappa$9

This yields a geometric explanation for the widely used score-difference-based detection metric.

Algorithmically, the workflow is concise. A sample is generated with DDIM and classifier-free guidance. At the final step $\partial_s\kappa$0, one computes either

$\partial_s\kappa$1

or

$\partial_s\kappa$2

The score-difference map is $\partial_s\kappa$3. The curvature-difference estimate uses Hutchinson: $\partial_s\kappa$4 with $\partial_s\kappa$5 random Rademacher vectors, followed by summation across channel indices for a spatial heatmap. The score-difference variant requires two forward/backwards per timestep and has negligible overhead. The curvature-difference variant requires $\partial_s\kappa$6 Hessian-vector products, approximately $\partial_s\kappa$7 additional backward passes. In practice $\partial_s\kappa$8–16, and even $\partial_s\kappa$9 suffices.

The empirical study uses Stable Diffusion v1.4 with baseline v1.1, v2.1 with baseline v2.0, and Realistic Vision v5.1, together with ground-truth template masks from Webster (2023), global-mem masks, and non-mem masks. Evaluation uses Intersection-over-Union and Pixel-Accuracy over $t$0 masks, with global detection by spatial averaging and AUC / TPR@1%FPR.

Method on SD v1.4	IoU	ACC
Bright Ending (template-verbatim only)	$t$1	$t$2
Raw curvature (template-verbatim only)	$t$3	$t$4
$t$5 with unconditional baseline (template-verbatim only)	$t$6	$t$7
$t$8 with unconditional baseline (template-verbatim only)	$t$9	$x_t\in\mathbb R^d$0
$x_t\in\mathbb R^d$1 on all cases	$x_t\in\mathbb R^d$2	$x_t\in\mathbb R^d$3
Bright Ending on all cases	$x_t\in\mathbb R^d$4	not reported

Aggregating $x_t\in\mathbb R^d$5 by spatial mean yields AUC $x_t\in\mathbb R^d$6 and TPR@1%FPR $x_t\in\mathbb R^d$7 for $x_t\in\mathbb R^d$8, far above raw curvature or Bright Ending. Heatmaps from $x_t\in\mathbb R^d$9 are sharply localized on memorized patches, whereas Bright Ending often lights up non-memorized regions. Later sampling steps, with $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$0 close to $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$1, dramatically improve the signal. Using SD v1.2 instead of v1.1 as baseline yields almost identical performance, and simple smoothing removes rare score-difference outliers.

The principal limitation is scope. The method is tailored to “verbatim” local memorization, specifically situations in which variability collapses coordinate-wise. It does not detect concept-level or style memorization that is globally distributed. It is also sensitive to late-timestep noise bounds and must be evaluated at sufficiently low noise, such as the final DDIM step.

3. Kink-robust curvature evaluation in level-set methods

In the level-set literature, curvature-difference methods arise from the failure of the textbook curvature formula when the signed-distance field contains kinks [(Lervåg et al., 2014); (Ervik et al., 2014)]. With interface $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$2 represented as the zero contour of $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$3, one writes

$$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$4

If $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$5 is reinitialized as a signed-distance function, then it is only $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$6 across points equidistant to two or more interfaces. In these regions, $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$7 has a jump. Standard second-order central differences cross the kink, sample non-smooth data, and cease to converge. The reported pathology includes curvature spikes of order $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$8, for example approximately $$s_\theta(x_t,c)\approx\nabla_{x_t}\log p_\theta(x_t\mid c), \qquad s_\theta(x_t)\approx\nabla_{x_t}\log p_\theta(x_t),$$9 on a mesh with $H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$0, and even spurious sign changes.

Three improved schemes were summarized as avoiding direct differentiation of $H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$1 across a kink. The Macklin–Lowengrub method fits a local quadratic interface parametrization by least squares, computes interface curvature from the smooth fitted polynomial, and extrapolates it back to nearby grid nodes. Lervåg’s method reconstructs a smooth local signed-distance field $H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$2 from a monotone cubic Hermite spline and then applies the standard CD-2 stencils to $H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$3. The Salac–Lu method keeps one level-set field per object, computes curvature on each smooth field separately, and blends them by

$H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$4

Method	Core mechanism	Reported outcome
MLM	Least-squares quadratic fit near interface	Curvature on contour itself; no differentiation across kink
LM	Spline-based local signed-distance reconstruction	Standard stencils applied to smooth $H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$5
SLM	Separate level sets per object and blend curvatures	Avoids medial-axis kink entirely
LOLEX	Local extraction, body separation, reinitialization, standard stencil	Robust in 2D and 3D near topological change

The benchmark behavior is consistent across test cases. For a static disk above a rectangle, CD-2 produces spikes up to $H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$6, while MLM, LM, and SLM all yield nearly uniform $H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$7 with pointwise errors $H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$8 or smaller. In a two-drop collision in shear flow, curvature spikes from CD-2 create artificial high pressure in the thin film and prevent coalescence, whereas MLM, LM, and SLM remove the spikes and recover coalescence at the correct time. MLM predicts slightly earlier coalescence, while LM and SLM agree almost exactly.

The LOcaL Level-set Extraction method extends the same logic by isolating a small local subgrid around problematic stencil points. Kink regions are detected by the quality function

$H_\theta(x_t,c)=\nabla_{x_t}s_\theta(x_t,c).$9

and whenever a local stencil contains both signs of $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$0 and $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$1, the algorithm copies $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$2 onto a small local block, identifies connected bodies by a bodyscan, extracts a separate $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$3 for each, extrapolates ghost cells, reinitializes by solving

$$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$4

and finally recomputes $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$5 and $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$6 on the smooth local field.

The quantitative validation is stronger than a mere artifact suppression claim. For a circle above a line with separations from $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$7 down to $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$8, the standard method saturates at $$c_i(x_t,c)=\bigl[\nabla^2_{x_t}\log p_\theta(x_t\mid c)\bigr]_{ii} =\bigl[-H_\theta(x_t,c)\bigr]_{ii},$$9 once the gap falls below approximately $i$0, whereas LOLEX stays within $i$1 of the analytical value $i$2 all the way to contact and converges with grid refinement. For concentric circles of width $i$3, LOLEX matches curve-fit normals to within max-RMS $i$4, while central and directional differences are off by $i$5 and $i$6 respectively. For a 3D sphere of radius $i$7 hovering $i$8 above a plane, the global $i$9-point stencil mis-estimates curvature by $\sigma^2$0 near the kink, whereas LOLEX reduces this to $\sigma^2$1.

The same pattern appears in two-phase flow simulations. Standard curvature produces unphysical pressure oscillations in thinning films, delays rupture, and yields interleaved spurious pressure patches just before film rupture and at neck maximum. LOLEX remains smooth and is reported to incur only a $\sigma^2$2 CPU overhead versus the standard curvature routine on a $\sigma^2$3 test. The main caveat is that a residual $\sigma^2$4–$\sigma^2$5 curvature error remains from local reinitialization, and the method is not conservative, with mass loss remaining $\sigma^2$6.

4. Finite-difference curvature quotients for geometric flows

In numerical geometric evolution, curvature-difference methods appear as stable discretizations of curvature-driven PDEs rather than as local interface diagnostics (Oberman et al., 2016, Deckelnick et al., 2022). Two distinct examples are prominent: affine-curvature flow in a level-set formulation and hyperbolic curvature flow for planar curves.

For affine-curvature flow, the governing PDE is

$\sigma^2$7

with Euclidean curvature

$\sigma^2$8

and equivalent form

$\sigma^2$9

A naive central-difference discretization is observed to blow up near points where $\kappa$00. The remedy is a monotone, viscosity-compatible scheme. One-sided differences define monotone approximations of $\kappa$01, while $\kappa$02 is replaced by a wide-stencil median formula

$\kappa$03

with neighbor values sampled on an approximately circular stencil of radius $\kappa$04. Since the cube-root map is not monotone, it is decomposed into nondecreasing pieces and regularized through

$\kappa$05

leading to an elliptic operator $\kappa$06 and a filtered scheme

$\kappa$07

Time stepping uses forward Euler with

$\kappa$08

where

$\kappa$09

The Barles–Souganidis framework then gives convergence to the unique viscosity solution. The filtered scheme attains nearly the accuracy of central differencing in smooth regions while remaining convergent.

For hyperbolic curvature flow in the plane, the continuous normal-flow equation is

$\kappa$10

A semidiscrete curve is represented by nodal values $\kappa$11 on a periodic grid $\kappa$12, with backward difference

$\kappa$13

The discrete curvature vector is approximated by

$\kappa$14

and the scalar curvature can be taken as

$\kappa$15

The resulting nodal ODE system is

$\kappa$16

The analysis proves a discrete energy law for $\kappa$17 and second-order convergence in $\kappa$18 under regularity assumptions and a uniform lower bound $\kappa$19: $\kappa$20 The fully discrete version uses $\kappa$21 with a semi-implicit update for the curvature term. Numerical experiments starting from a smooth strictly convex ellipse show bounded curvature initially, followed by finite-time blow-up of

$\kappa$22

and development of two corners. Outward normal initial velocity delays blow-up, and damping $\kappa$23 postpones the singularity further.

Taken together, these two lines of work show distinct operational meanings of a curvature-difference method. In affine flow, the emphasis is monotonicity, consistency, and explicit-Euler stability for a degenerate nonlinear PDE. In hyperbolic curvature flow, the emphasis is a direct curvature difference quotient on a moving mesh, coupled to energy control and second-order convergence.

5. Adaptive redistribution and tangential-velocity formulations for Willmore flow

For Willmore flow, curvature-difference methods are embedded in adaptivity rather than only in force evaluation (Duan et al., 4 Jan 2026). Let $\kappa$24 be a smooth closed planar curve with parametrization $\kappa$25 and arc-length derivative $\kappa$26. The scalar curvature and its variation are

$\kappa$27

The Willmore energy is

$\kappa$28

and the $\kappa$29-gradient flow satisfies

$\kappa$30

Time discretization uses BDF$\kappa$31: $\kappa$32 with BDF1 coefficients $\kappa$33, BDF2 coefficients $\kappa$34, and BDF3 coefficients $\kappa$35. The fully discrete finite-difference method couples the BDF approximation to centered spatial differences for $\kappa$36, $\kappa$37, and $\kappa$38. Nonlinearity is treated by Picard or Newton–Picard iteration. No explicit CFL-type constraint is reported beyond choosing $\kappa$39 small enough for convergence of the nonlinear solver.

The first adaptive strategy is weighted-arc-length equidistribution, denoted A-WAR. A positive scalar monitor $\kappa$40 encodes $\kappa$41 and $\kappa$42, and the weighted arc-length density is

$\kappa$43

The equidistribution principle requires

$\kappa$44

Discrete weights are

$\kappa$45

Adaptive monitor selection uses

$\kappa$46

with monitor choices switching among

$\kappa$47

and, when curvature variation is large,

$\kappa$48

The second strategy, A-BDF$\kappa$49-FDM, removes explicit redistribution by adding a tangential velocity enforcing

$\kappa$50

Starting from the mesh functional

$\kappa$51

the Euler–Lagrange equation gives

$\kappa$52

A relaxation dynamics for $\kappa$53 then yields

$\kappa$54

with tangential velocity

$\kappa$55

An additional Energy-Stable Correction algorithm introduces

$\kappa$56

and an auxiliary variable $\kappa$57. After a provisional update, one computes

$\kappa$58

so that

$\kappa$59

This guarantees discrete energy decay at the theoretical level.

The reported numerical example uses the oscillatory initial curve

$\kappa$60

$\kappa$61

Standard BDF1-FDM leads to mesh ratio

$\kappa$62

growing to $\kappa$63 by $\kappa$64, with clustering and eventual breakdown. By contrast, A-WAR and A-BDF1-FDM keep both $\kappa$65 and the weighted ratio

$\kappa$66

below $\kappa$67 for all $\kappa$68, while accurately capturing the evolving curve without mesh tangling.

6. Curvature estimation from volume fractions and comparative trade-offs

Curvature estimation from volume fractions provides a useful comparative frame because it isolates the accuracy-cost structure of several interface-curvature estimators on Cartesian meshes (Han et al., 2023). Four methods are compared: the height function (HF), PLIC-centroidal fitting (PC), PLIC-volumetric fitting (PV), and volumetric fitting (VF).

HF defines a pseudo-normal direction from the largest component of $\kappa$69, forms column sums of the volume fraction $\kappa$70, and computes derivatives of the resulting height field by second-order central differences. The mean curvature estimate is

$\kappa$71

For exact $\kappa$72, HF is $\kappa$73 in curvature and converges to machine precision in static random paraboloid tests, but it fails more often when $\kappa$74 because complete columns are unavailable. If $\kappa$75 comes from second-order VOF advection and is only $\kappa$76 in $\kappa$77, the curvature may degrade to $\kappa$78 for small $\kappa$79.

PC and PV fit a local paraboloid

$\kappa$80

in a rotated frame aligned with the target-cell normal, then evaluate curvature from the fitted coefficients. PC minimizes a weighted least-squares error over neighboring PLIC centroids, using radial and area-projection weights. PV instead matches the actual PLIC planes in a least-squares sense of projected area and assembles a symmetric $\kappa$81 normal system from polygon moments computed by Green’s theorem. VF fits the same paraboloid directly to the volume fraction field by minimizing

$\kappa$82

using exact analytic polyhedron-paraboloid intersections and a Levenberg–Marquardt solve.

Method	Static behavior	Dynamic / noisy behavior	Cost
HF	$\kappa$83 to machine precision	Degrades for advected $\kappa$84; column failures for high curvature	$\kappa$85
PC	$\kappa$86 at low $\kappa$87	Highest spurious currents of the three dynamic methods	$\kappa$88
PV	$\kappa$89 only at very low $\kappa$90	Best balance; lowest overall spurious current in 3D translating droplet	$\kappa$91
VF	$\kappa$92 to machine precision	Dynamic tests not performed	$\kappa$93

The detailed results sharpen these trade-offs. In static random paraboloids, HF and VF both reach second-order convergence down to machine precision. PC and PV are second-order down to approximately $\kappa$94, after which they transition to $\kappa$95 behavior. Under perturbations of $\kappa$96 with $\kappa$97, PV maintains lower error than HF and PC across all $\kappa$98, while VF errors are less than or approximately equal to PV errors. Increasing the PV stencil from $\kappa$99 to $\partial_s\kappa$00 extends convergence down to $\partial_s\kappa$01, with only minor further gain at $\partial_s\kappa$02.

In dynamic tests, HF gives the lowest spurious velocities for a static droplet at high Laplace number, but in a 3D translating droplet it produces higher capillary numbers when $\partial_s\kappa$03 has large transport errors. PV gives the lowest $\partial_s\kappa$04 among the three tested dynamic methods and is recommended as the default curvature estimator when coupling to a two-phase Navier–Stokes solver. HF remains attractive as a first pass, with PV as a backup where height columns fail. VF is principally reserved for static or near-static configurations where extremely high-fidelity curvature is required and its two-orders-of-magnitude cost penalty is acceptable.

A broader implication, suggested by the combined literature, is that curvature-difference methods are best understood not as a single numerical recipe but as a design principle. They intervene precisely where naive curvature evaluation is confounded: by intrinsic data rigidity in diffusion models, by non-differentiability in signed-distance fields, by loss of monotonicity in nonlinear PDE discretizations, by mesh distortion in fourth-order geometric flows, or by transport noise in VOF reconstructions. The recurring result is not that curvature becomes universally easy to compute, but that reliable curvature information depends on explicitly controlling the mechanism that corrupts it.