---
title: 'Spaghetti Splatting: Concepts & Applicants'
url: https://www.emergentmind.com/topics/spaghetti-splatting
type: topic
---

# Spaghetti Splatting: Concepts & Applicants

Spaghetti Splatting denotes two distinct research concepts. In magnetic materials, it is an interpretive description of rotating a disordered “helical spaghetti” state into an oriented skyrmion lattice; the terminology of the underlying work is “chiral jammed state,” “helical spaghetti,” and “precipitating ordered skyrmion lattices” [1810.07631]. In differentiable rendering, “Spaghetti Splatting” is the name of a representation in which connected line segments or polylines are convolved with spatial kernels and rendered by volume integration, extending point-based splatting to curvilinear geometry [2611.12716]. The shared metaphor is the transformation of spatially distributed, elongated structures into an organized representation, but the physical mechanisms and mathematical frameworks are unrelated.

## 1. Terminological scope and conceptual distinction

The magnetic usage describes a jammed mixture of chiral magnetic textures subjected to a symmetry-breaking rotation of a sample in a static magnetic field. The disordered state may contain randomly oriented helices, labyrinth domains, rotationally disordered skyrmion lattices, isolated skyrmions, and other imperfectly terminated chiral structures. Rotation through particular field orientations disentangles this state and promotes an oriented hexagonal skyrmion lattice.

The rendering usage concerns differentiable image synthesis. Its primitive is a finite line segment rather than an isolated point, and connected collections of segments form polylines. Each geometric measure is convolved with a kernel, such as a Gaussian or compact Beta kernel, to produce a volumetric density that can be integrated along camera rays. The resulting representation is optimized from images by gradient descent and is intended for hair, fibers, wires, strokes, bristles, and other elongated structures.

The two usages should not be conflated. Magnetic “splatting” is a metaphor for field-induced reorganization of a chiral texture, whereas rendering Spaghetti Splatting is a formal differentiable primitive based on convolution, volume rendering, and front-to-back compositing. The common term refers to the visual or structural analogy of disordered elongated components being reorganized into coherent strands or lattices.

## 2. Magnetic helical spaghetti and skyrmion precipitation

In the magnetic context, “spaghetti” is not a single thermodynamic phase. It denotes a topologically and geometrically frustrated mesh of chiral structures with broadly similar characteristic length scales but without common long-range orientation. In chemically disordered $(\mathrm{Fe},\mathrm{Co})\mathrm{Si}$, random occupation of the Fe/Co sublattice produces spatial variations in exchange, Dzyaloshinskii–Moriya interaction (DMI), and magnetocrystalline anisotropy. These variations provide nucleation and pinning sites.

A randomly oriented population of helices is characterized by propagation vectors $\mathbf q$ distributed over many directions. In reciprocal space, one ideal helix produces two peaks at $\pm\mathbf q$, whereas a population with many propagation directions produces a ring. Labyrinth domains, consisting of elongated winding magnetic domains, also generate ring-like scattering. Rotationally disordered skyrmion lattices retain a characteristic periodicity while lacking a common azimuthal orientation and consequently contribute to a ring. Isolated or distorted skyrmions can arise when domains nucleate at defects, collide, or terminate without matching an integer number of helix or skyrmion periods.

The term “jammed” refers to the inability of neighboring domains to expand, annihilate, or rotate past one another. A chiral domain can grow until it encounters another domain or a defect; misoriented neighbors then become trapped. The resulting state is a mixture of helices, labyrinths, skyrmion domains, and related textures rather than a uniform phase.

“Precipitation” in this setting does not mean chemical precipitation. It describes the selection, growth, and separation of an ordered skyrmion phase from the pre-existing disordered mixture by rotating the sample in a static field. The principal experimental result is that sufficiently large or repeated rotations through particular field orientations convert a magnetically disordered state into an oriented hexagonal skyrmion lattice, even in chemically disordered crystals, weak-anisotropy systems, powders, and polycrystalline materials [1810.07631].

For a two-dimensional unit magnetization field $\mathbf m(x,y)$, the skyrmion number is conventionally written as

$$
N_{\mathrm{sk}} =
\frac{1}{4\pi}
\int
\mathbf m\cdot
\left(
\frac{\partial \mathbf m}{\partial x}
\times
\frac{\partial \mathbf m}{\partial y}
\right)
\,dx\,dy .
$$

A simple in-plane rotation of a skyrmion tube preserves its topological charge. By contrast, the conversion of labyrinth domains into closed skyrmions changes the number of topological objects. The reported simulations associate this restructuring with zero-moment monopoles or Bloch-point monopoles: labyrinth domains break into smaller closed objects, skyrmions can split along their length, and skyrmions can detach from grain surfaces through monopole propagation along the tube before annihilation.

## 3. Field-induced reorientation and scattering signatures

Small-angle neutron scattering (SANS) distinguishes the magnetic configurations through reciprocal-space periodicity and angular organization. A uniform ring indicates a characteristic period without a net in-plane orientation. Six sharp or broadened peaks indicate a hexagonal skyrmion lattice with a common orientation. Angular disorder broadens the six peaks azimuthally, while period disorder broadens them radially.

For the $(\mathrm{Fe},\mathrm{Co})\mathrm{Si}$ measurements, ring and six-fold features were centered at approximately

$$
q=0.164\pm0.005\ \mathrm{nm}^{-1},
$$

corresponding to a real-space period of approximately

$$
\frac{2\pi}{q}\approx77\ \mathrm{nm}.
$$

For powdered $\mathrm{Cu_2OSeO_3}$, the initial ring radius was approximately $q\approx0.103\ \mathrm{nm}^{-1}$, corresponding to a period of about $61$ nm. The similar radial positions before and after treatment show that rotation primarily changes orientation and phase content rather than the local pitch.

For a sample initially having $H\parallel[100]$, the rotation sequence exhibits three characteristic stages:

1. **Initial rotation toward the hard-axis vicinity**: As the field rotates from $[100]$ toward $[110]$, six-fold intensity increases, azimuthal peak width decreases, and the ring remains nearly constant. The oriented-skyrmion fraction changes only slightly, from about $50\%$ to $54\%$. This indicates alignment of pre-existing skyrmion domains rather than immediate conversion of the ring phase.

2. **Passage beyond the hard-axis vicinity**: The ring intensity decreases while the six-fold amplitude remains relatively constant. Disordered helical and labyrinth structures are destroyed or reorganized, and the oriented-skyrmion fraction increases substantially. This is the stage most directly associated with de-jamming and skyrmion formation.

3. **Return to the original field direction**: The sample does not recover its initial SANS pattern. Six-fold scattering becomes stronger, the ring is reduced, and the oriented-skyrmion fraction reaches approximately $80\%$. The transformation is therefore irreversible under the treatment.

Starting from $H\parallel[110]$ produces a similar effect: after a $180^\circ$ rotation, the skyrmion fraction increases from approximately $37\%$ to $79\%$. Starting from $H\parallel[111]$ and rotating in the $(110)$ plane also transforms the ring into six-fold scattering. The final lattice orientation is the same as that obtained from rotation in the $(001)$ plane, indicating that rotation direction and dynamic treatment influence the final orientation more strongly than the initial crystal-axis orientation.

In powdered $\mathrm{Cu_2OSeO_3}$, the initial pattern consists only of a ring because the crystallites are misoriented. Repeated $90^\circ$ rotations, reaching $\phi=1440^\circ$, progressively produce six-fold order. The ordered component accounts for $16\pm0.7\%$ of the total sample scattering. The lower ordered fraction is consistent with weak intergrain coupling and only a subset of crystallites becoming aligned with the treatment-induced orientation.

## 4. Energetic and micromagnetic mechanism

The micromagnetic energy contains exchange, DMI, Zeeman, magnetocrystalline-anisotropy, and magnetostatic or demagnetizing contributions. Schematically,

$$
E=
\int_V
\left[
A(\nabla\mathbf m)^2
+
D\,\mathbf m\cdot(\nabla\times\mathbf m)
-
\mu_0M_s\mathbf H\cdot\mathbf m
+
E_{\mathrm{ani}}
+
E_{\mathrm{demag}}
\right]\,dV .
$$

The Zeeman term favors moments parallel to the field,

$$
E_Z=-\mu_0M_s\int_V \mathbf H\cdot\mathbf m\,dV,
$$

whereas the DMI favors twisting,

$$
E_{\mathrm{DMI}}
=
D\int_V
\mathbf m\cdot(\nabla\times\mathbf m)\,dV .
$$

The reported mechanism is dynamic rather than a simple relaxation into a pre-existing static minimum. As the field direction approaches a magnetic hard axis, the moments cannot remain everywhere field-parallel while preserving the preferred chiral texture. They increasingly diverge away from the applied field, producing larger transverse components and texture distortions. Exchange energy rises, while the DMI contribution becomes more negative. Near the hard-axis passage, the skyrmion lattice reorients; after the passage, it settles into an orientation determined by the rotation geometry, skyrmion–skyrmion repulsion, and grain shape.

The simulations show energy-curve inflections near $45^\circ$, associated with the relevant hard-axis passage. Anisotropy energy remains close to zero or nearly constant in the simulated cases, and reorientation persists when the anisotropy is weak or zero. Magnetocrystalline anisotropy therefore helps define local energy landscapes but is not by itself necessary for the rotational ordering mechanism.

Micromagnetic simulations use OOMMF with a $3$-nm cubic mesh, cylindrical and parallelepiped grains, finite-length skyrmion tubes, and random cubic anisotropy strengths between $K_U=0$ and $50\times10^3\ \mathrm{J\,m^{-3}}$. Two initially different lattice orientations remain distinct up to approximately $\phi=34^\circ$, begin to converge thereafter, and reorient near $\phi=45^\circ$. In a cylindrical grain, the common orientation is approximately $\theta\approx0^\circ$; in a parallelepiped, it is approximately $\theta\approx30^\circ$. Geometry therefore affects the final angular offset, but convergence toward a common orientation after rotation is robust and largely independent of the initial state.

Repeated hard-axis passages act as annealing-like events. Field rotation overcomes pinning, forces domains to move and interact, aligns existing skyrmions, breaks up disordered helices and labyrinths, promotes new skyrmion formation, and reinforces the selected lattice orientation. The resulting orientation is not necessarily determined by the initial lattice orientation or by a single globally dominant crystalline easy axis.

## 5. Convolutional curve representation in differentiable rendering

Rendering Spaghetti Splatting generalizes splatting through two independent choices: the smoothing kernel and the geometric support measure. Point-based splatting uses a point measure convolved with a kernel, whereas Spaghetti Splatting replaces the point measure with a segment or polyline measure while retaining convolution and volume rendering [2611.12716].

Let a continuous curve be

$$
\gamma:[0,1]\rightarrow\mathbb R^3 .
$$

Its parameterized curve measure is

$$
m_\gamma(\mathbf x)
=
\int_0^1
\delta\!\left(\mathbf x-\gamma(u)\right)\,du .
$$

Convolution with a kernel $K$ gives

$$
(m_\gamma*K)(\mathbf x)
=
\int_0^1
K\!\left(\mathbf x-\gamma(u)\right)\,du .
$$

For a finite segment with endpoints $\mathbf a$ and $\mathbf b$, define $\mathbf v=\mathbf b-\mathbf a$ and

$$
\mathbf x(u)=\mathbf a+u\mathbf v,\qquad u\in[0,1].
$$

The segment primitive is

$$
P(\mathbf x)
=
\int_0^1
K\!\left(\mathbf x-\mathbf a-u\mathbf v\right)\,du .
$$

A polyline with vertices $\mathbf p_0,\ldots,\mathbf p_M$ consists of connected segments. Its measure is the sum of the segment measures, and its convolved field is the corresponding sum of kernel responses.

The representation has a strict point-based limit. When $\mathbf a=\mathbf b$, the segment becomes a point measure, and convolution with an anisotropic Gaussian produces an anisotropic 3D Gaussian primitive. Nonzero-length segments and polylines add explicit elongated support, connectivity, tangent structure, and ordered control points that cannot generally be represented by a single point Gaussian.

Polyline approximations converge to the convolved continuous curve when the polyline converges uniformly to the curve and the kernel is continuous, bounded, and sufficiently regular. If $\gamma_h\to\gamma$ uniformly, then

$$
P_h(\mathbf x)\longrightarrow P(\mathbf x)
$$

as the maximum segment length or parameter spacing $h$ tends to zero. For a kernel with Lipschitz constant $L_K$,

$$
\left|P_h(\mathbf x)-P(\mathbf x)\right|
\le
L_K\left\|\gamma_h-\gamma\right\|_\infty .
$$

This convergence concerns the convolved field and rendered result; it does not guarantee recovery of the semantic identity or topology of physical strands.

## 6. Rendering, optimization, and applications

For a Gaussian kernel with covariance $\Sigma$, the segment density is

$$
\sigma(\mathbf x)
=
\rho
\int_0^1
N_\Sigma\!\left(\mathbf x-\mathbf x(u)\right)\,du ,
$$

where $\rho$ is a learned density or opacity scale. Along a camera ray $\mathbf r(t)=\mathbf o+t\mathbf d$, the segment optical thickness is

$$
\tau(\mathbf r)
=
\int_{-\infty}^{\infty}
\sigma(\mathbf r(t))\,dt .
$$

The ray-integrated Gaussian reduces to a one-dimensional Gaussian in the segment coordinate. With

$$
Q=\frac12\Sigma^{-1},
$$

the projected precision is

$$
Q_\perp
=
Q-
\frac{Q\mathbf d\mathbf d^\top Q}
{\mathbf d^\top Q\mathbf d},
$$

and the segment-coordinate precision is

$$
\kappa^2=\mathbf v^\top Q_\perp\mathbf v .
$$

The remaining segment integral has an error-function closed form. Opacity is obtained from

$$
\alpha(\mathbf r)=1-\exp\!\left(-\tau(\mathbf r)\right).
$$

Connected segments are ordered and composited front to back using volume-rendering transmittance,

$$
\mathbf C
=
\sum_k T_k\alpha_k\mathbf c_k,
\qquad
T_k=\prod_{j<k}(1-\alpha_j).
$$

Because a shared endpoint has measure zero under the one-dimensional segment integral, connected segments do not double-count a finite density contribution at their joints. Correct optical-thickness composition also makes subdivision compatible with the parent segment when density parameters are distributed consistently.

The rendering system locally linearizes camera projection around each segment midpoint. A projection $\pi$ is approximated as

$$
\pi(\mathbf x)
\approx
\pi(\mathbf m)
+
J_\pi(\mathbf m)(\mathbf x-\mathbf m),
$$

where $\mathbf m$ is the segment midpoint and $J_\pi(\mathbf m)$ is the projection Jacobian. The resulting screen-space footprint supports EWA-style rasterization, foreshortening, tile-based splatting, GPU rendering, and anti-aliasing. Gaussian kernels provide error-function-based ray integration; compact Beta kernels provide finite spatial support. The geometric measure and kernel are independent design choices.

Training is differentiable with respect to endpoints, covariance or width, density, opacity, color, kernel parameters, and appearance parameters. The loss combines reconstruction, MCMC regularization, and connectivity terms. Population control splits long segments at their midpoints, clones short segments, relocates or prunes low-opacity segments, merges nearby polyline endpoints, and applies endpoint attraction regularization. These operations serve different purposes: splitting improves local curvature and depth representation, cloning redistributes capacity, and merging and attraction increase connected-component size even when their direct effect on photometric metrics is limited.

The method is evaluated on general scenes, hair, fibers, and other line-like structures. On general scenes, Gaussian and Beta variants remain competitive with point-based splatting. Beta Spaghetti Splatting performs best on Mip-NeRF 360 in the reported comparison, while the Gaussian variant performs best on Shelly and Fuzzy. The method substantially outperforms the opaque Line Primitives baseline on general scenes, whose constant screen-space width is unsuitable for broad surfaces, perspective-dependent geometry, and depth variation.

For hair reconstruction, experiments use eight GeomHair hairstyles and a challenging curly-hair model. The Gaussian variant reports $32.25$ PSNR, $0.940$ SSIM, $0.043$ LPIPS, and F1 scores of $61.2$, $80.3$, and $88.9$ at $2$, $3$, and $4$ mm, respectively, compared with $29.56$ PSNR, $0.923$ SSIM, $0.064$ LPIPS, and F1 scores of $46.9$, $63.5$, and $72.7$ for Gaussian Haircut. The paper summarizes these results as a $2.7$ dB PSNR improvement and a $14.3$-point improvement in F1 at the $2$ mm threshold.

The explicit polyline output is directly amenable to position-based dynamics and related strand simulation systems. It provides ordered control points, tangent directions, resampling operations, connected components, and a structure suitable for distance and bending constraints. This distinguishes the representation from a point cloud whose connectivity must be inferred after reconstruction.

Important limitations remain. High-frequency appearance may require substantial geometric subdivision or additional learnable texture representations. Local projection linearization can become inaccurate for long segments or strong depth variation. Connectivity is encouraged but not guaranteed, and the method does not establish automatic recovery of true physical topology. Hair remains difficult because of occlusion, self-similarity, complex shading, and image ambiguity. The current construction does not automatically recover physical properties such as stiffness, damping, mass, or material parameters, and the experiments use per-scene optimization rather than feedforward reconstruction.

The two meanings of Spaghetti Splatting therefore converge only at the level of metaphor. In magnetic materials, rotational forcing reorganizes a jammed mixture of chiral textures into a collectively oriented skyrmion lattice. In differentiable rendering, convolution transforms connected curvilinear measures into smooth volumetric primitives that can be optimized and rendered efficiently. Both descriptions concern the emergence of ordered structure from distributed elongated components, but one is a field-driven micromagnetic phenomenon and the other is a kernel-based representation for inverse graphics.

Source: https://www.emergentmind.com/topics/spaghetti-splatting