Papers
Topics
Authors
Recent
Search
2000 character limit reached

Spaghetti Splatting: Differentiating Curvilinear Dirac Masses via Convolution

Published 9 Nov 2026 in cs.CV and cs.GR | (2611.12716v1)

Abstract: In production systems, 3D geometry is represented by primitives such as points, lines, and triangles, as rendering hardware is optimized to generate images from them very efficiently. Learning such primitives from images by volume rendering is challenging, as they are Dirac masses defined at infinitesimal locations in space. EWA splatting, Poisson reconstruction, and 3D Gaussian Splatting overcome this by convolving Dirac masses with a low-frequency kernel, a recipe that generalizes along two axes: the kernel and the primitive. As points and triangles have already received significant attention, we investigate this recipe for curvilinear primitives, and introduce Spaghetti Splatting. We approximate curves by polylines whose convolution converges to that of the continuous curve, and volume render each segment so that connected segments composite without overcounting at their joints. For fast training and inference, we linearize the camera projection per segment, which also accommodates kernels beyond Gaussians, such as Beta kernels. As a point is a curve of zero length, our representation is a strict superset of 3D Gaussian Splatting. It remains competitive with the state-of-the-art on general scenes, outperforms methods specialized to line-like geometry, natively supports anti-aliasing, and produces long connected strands that are directly amenable to physics simulation.

Summary

  • The paper generalizes differentiable image rendering by replacing point primitives with finite line segments and connected polylines.
  • A closed-form expression for optical depth is derived for the segments, allowing real-time rendering and differentitional properties without sampling points.
  • Evaluations demonstrate Spaghetti Splatting exceeds the image reconstruction quality of existing point-based methods for various challenging scenes and straps within elongated structures, such as hairs.

Problem formulation and central contribution

Spaghetti Splatting addresses the mismatch between explicit curvilinear geometry and differentiable image-based reconstruction. Conventional rendering primitives—points, lines, and triangles—are efficient for rasterization, but their idealized geometric measures are singular. Directly optimizing such measures through volume rendering is poorly conditioned because infinitesimal support provides no volumetric neighborhood on which image-space gradients can act. 3D Gaussian Splatting resolves this issue for point primitives by convolving Dirac masses with anisotropic Gaussian kernels (Kerbl et al., 2023). The paper generalizes the same construction along a second axis: instead of changing only the kernel, it replaces the point support with finite line segments and connected polylines.

The resulting representation consists of learnable polylines convolved with Gaussian or compact Beta kernels. Each segment contributes a volumetric density, and neighboring segments are composited independently along camera rays. A zero-length segment recovers a point primitive, making the representation a strict superset of anisotropic point splatting rather than a competing parameterization with disjoint expressivity. This design is important because it permits a single representation to model both localized volumetric structure and elongated, connected geometry.

The paper’s conceptual decomposition is particularly useful. A differentiable primitive is characterized by two independent choices: the support measure and the convolution kernel. A point measure convolved with a Gaussian yields a 3DGS-like primitive; a point measure convolved with a Beta kernel yields a DBS-like primitive (Sun et al., 15 Mar 2025). Replacing the point measure with a segment produces an extended primitive, and concatenating segments produces a curvilinear primitive.

Figure 1

Figure 1: The representation generalizes splatting by independently varying the support measure and convolution kernel.

This formulation has a direct geometric implication: the learned representation is not merely a collection of nearby points approximating a curve. It explicitly stores connected segments, which can preserve strand continuity and can be passed to downstream geometric procedures such as position-based dynamics. The authors nevertheless qualify broader topological claims: connectivity is encouraged algorithmically, but automatic recovery of semantically correct topology is not established in general.

Convolutional segment primitives

A segment with endpoints aa and bb is parameterized as x(u)=a+u(b−a)x(u)=a+u(b-a) for u∈[0,1]u\in[0,1]. Its support is represented as a one-dimensional measure embedded in R3\mathbb{R}^3. Convolution with an anisotropic Gaussian produces a density field obtained by integrating the Gaussian response over the segment. For a point query xx, the density is therefore the accumulated response of every point on the segment, rather than the response of a single center.

This construction retains the principal optimization advantages of Gaussian splatting while changing the support geometry. In particular, the segment length and orientation become differentiable parameters. When the endpoints coincide, the segment measure degenerates to a point mass and the conventional point-based primitive is recovered. The same derivation extends to compact Beta kernels, although the corresponding expressions differ from the Gaussian case.

The paper establishes convergence from polylines to continuous curves. As the segment length decreases, independently rasterized segments converge to the convolution of the underlying smooth curve. The experiments visualize this behavior using polylines with 4, 8, 16, and 32 Beta segments. The absence of visible knots at segment joints is significant: independent segment evaluation does not inherently introduce discontinuities in the rendered density when the segments share endpoints and are composited correctly.

Figure 2

Figure 2: Increasing the number of segments produces a smooth rendered curve without visible artifacts at the joints.

The convergence claim concerns the convolved volumetric representation and its rendering, not exact recovery of a latent continuous curve from finite images. The latter remains dependent on camera coverage, kernel bandwidth, regularization, and the capacity-control procedure used during optimization.

Closed-form ray integration and compositing

The central technical result is a closed-form expression for the optical depth contributed by a Gaussian-convolved segment. Substituting the segment convolution into the volume-rendering integral yields a double integral over the camera ray and the segment parameter. Since the Gaussian density is nonnegative, the order of integration can be exchanged. Integrating the Gaussian along the ray reduces the remaining segment-domain function to a one-dimensional Gaussian in uu.

Consequently, the segment optical depth can be evaluated using the error function rather than numerical quadrature. The expression depends on the ray direction, the segment endpoints, the anisotropic covariance, and the closest-response location along the segment. This closed form is the basis for differentiating both geometric and appearance parameters without sampling the segment or ray at multiple points.

The treatment of connected polylines is equally important. Each segment is evaluated as an individual primitive and then composited front-to-back. Because the opacity is derived from integrated density, adjacent segments do not double-count their shared endpoint in the manner that naïve point sampling could. The authors further observe that subdividing a segment into children with inherited density leaves the image-space rendering exactly unchanged. This property supports adaptive refinement without introducing a representation-dependent photometric discontinuity.

The implementation does not evaluate the full ray-dependent closed form independently for every primitive-pixel pair. Instead, it applies an EWA-style screen-space linearization at the segment midpoint and scales the projected kernel according to inverse depth. This approximation reduces the computation to a small set of screen-space quantities while preserving depth-dependent apparent width. It also allows the same rendering framework to support Gaussian and Beta profiles.

The resulting method is therefore a hybrid of analytic volume rendering and splatting-style rasterization. The analytic derivation supplies differentiability and correct segment integration; the screen-space approximation supplies the throughput required for per-scene optimization and interactive inference.

Optimization, anti-aliasing, and connectivity

The training objective combines a photometric reconstruction loss with the opacity and width regularization used by 3DGS-MCMC (Kheradmand et al., 2024) and an attraction term for nearby polyline endpoints. The endpoint attraction is not required for image reconstruction in the narrow sense: the ablation study shows that disabling it has little direct effect on PSNR. Its role is structural, improving the probability that independently initialized segments merge into cohesive connected components.

Population control adapts the MCMC relocation strategy to segments. Low-opacity segments are pruned or relocated, long segments are split at their midpoint, and short segments are cloned. These operations serve different purposes. Midpoint splitting increases local geometric resolution while preserving the parent’s rendering; cloning redistributes representation capacity to regions that require additional short segments. The ablations show that removing cloning produces a substantially larger degradation than removing midpoint splitting, indicating that capacity relocation is more important than subdivision alone.

The structural effects are visible in the reported ablation results:

Removed component PSNR Mean connected-component size
Midpoint splitting 33.13 3.26
Segment cloning 32.54 4.00
Polyline merging 33.59 1.00
Attraction regularization 33.57 2.77
None 33.54 3.50

The ablation supports a distinction between appearance quality and geometric organization. Polyline merging and attraction regularization do not improve reconstruction metrics directly, but removing merging collapses the mean component size to 1.00, while removing attraction reduces structural cohesion. Thus, connectedness is an emergent property of the optimization machinery rather than a consequence of the rendering equation alone.

Anti-aliasing follows Mip-Splatting (Yu et al., 2023). A three-dimensional filter limits the frequency content of primitives to what the training views can resolve, while a two-dimensional screen-space filter suppresses subpixel flickering during rasterization. Under the local projection model, filtering amounts to adding covariance terms to the projected kernel, so the closed-form segment integral remains applicable. This is a substantive advantage over opaque line primitives that lack spatial convolution and require multisample anti-aliasing.

Evaluation on general scenes

The general-scene evaluation covers 23 scenes from Mip-NeRF 360 (Etemadi et al., 2021), Shelly, and the concurrent Fuzzy benchmark for fibrous objects. The baselines include 3DGS-MCMC and Deformable Beta Splatting, both coupled with an eight-site Spherical Voronoi appearance model [2610.???], as well as opaque Line Primitives [2611.???]. The experimental protocol uses 30,000 optimization iterations, scene-specific primitive budgets, and held-out views.

The results show that the relative advantage of the two kernels depends on the dataset. On Mip-NeRF 360, the Beta version obtains the best reported PSNR, SSIM, and LPIPS among the listed methods: 28.78 dB, 0.838, and 0.193, respectively. On Shelly, the Gaussian version performs best, reaching 37.38 dB PSNR, 0.962 SSIM, and 0.048 LPIPS. On Fuzzy, the Gaussian version again leads PSNR and SSIM at 35.03 dB and 0.957, while the Gaussian and Beta variants tie the best LPIPS at 0.040.

Dataset Best variant PSNR SSIM LPIPS FPS
Mip-NeRF 360 Beta 28.78 0.838 0.193 129
Shelly Gaussian 37.38 0.962 0.048 452
Fuzzy Gaussian 35.03 0.957 0.040 415

The method remains competitive in rendering speed. Its reported rates range from 129 to 415 FPS across the three benchmarks, comparable to 3DGS-MCMC and substantially faster than many volumetric alternatives. Opaque Line Primitives are faster on these measurements, reaching 287, 731, and 858 FPS, but their constant screen-space width produces a severe representational limitation. They obtain only 8.66 dB PSNR on Mip-NeRF 360, compared with 28.78 dB for the proposed Beta representation. This contrast supports the paper’s claim that line-like primitives must retain correct depth-dependent volumetric behavior to operate on general scenes.

Figure 3

Figure 3: Spaghetti Splatting combines thin-structure reconstruction with the ability to model broad, textureless regions through variable-width kernels.

The qualitative comparisons explain the numerical pattern. Point-based methods can approximate thin elongated geometry, but they do so with scattered supports that do not form continuous strands. The proposed representation follows elongated structures with connected segments while retaining the ability to use wide kernels for surfaces and shadows. Conversely, the concurrent line representation is specialized to constant-width line geometry and is therefore poorly matched to scenes containing both thin structures and spatially broad appearance.

Hair reconstruction and geometric fidelity

Hair reconstruction provides a more direct test of the representation’s curvilinear inductive bias. The authors evaluate eight hairstyles from GeomHair and an additional curly hair model, rendering 100 views per hairstyle and holding out every eighth view. The comparison is against Gaussian Haircut (Zammit-Mangion et al., 2024), which incorporates hair-specific priors but represents the result with oriented Gaussians rather than explicit connected strands.

The proposed Gaussian-kernel model achieves 32.25 dB PSNR, 0.940 SSIM, and 0.043 LPIPS, compared with 29.56 dB, 0.923, and 0.064 for Gaussian Haircut. The geometric F1 scores are 61.2, 80.3, and 88.9 at 2, 3, and 4 mm, respectively, compared with 46.9, 63.5, and 72.7. Thus, the reported improvement is 2.69 dB in PSNR and 14.3 percentage points in F1 at the strictest 2 mm threshold.

Figure 4

Figure 4: Explicit polyline strands improve both novel-view reconstruction and strand-level geometric correspondence relative to Gaussian Haircut.

The significance of this result is not limited to image quality. A collection of connected polylines is directly compatible with strand-level operations and physics simulation, whereas a cloud of anisotropic Gaussians requires an additional geometric extraction step. The result also demonstrates that explicit strand structure can be learned without a dedicated learned hair prior. However, the evaluation uses rendered views, selected hairstyles, and a controlled comparison protocol; it does not establish comparable performance on unconstrained real-world hair captures.

Limitations and open questions

The representation has an unfavorable capacity trade-off for high-frequency texture. Because appearance variation is encoded through primitives rather than an independent texture representation, fine texture may require many short segments. This can increase memory use, complicate connectivity, and weaken the geometric interpretation of the recovered polylines. The paper identifies learnable texture maps as one possible way to decouple appearance frequency from geometric frequency, but does not evaluate that design.

The screen-space renderer also relies on local camera-projection linearization and midpoint EWA evaluation. The reported results demonstrate that this approximation is effective in the tested regimes, but the manuscript does not provide a complete characterization of its error under extreme foreshortening, highly curved segments, large kernels, or strong perspective variation within a segment.

The topology induced by endpoint merging and attraction regularization is likewise heuristic. The method produces coherent strands, but the ablations show that connectivity is sensitive to optimization choices. It remains open whether the procedure can reliably recover branching, crossing, termination, and occluded topology without producing false joins or disconnected fragments.

Finally, the paper evaluates Gaussian and Beta kernels but does not investigate kernel interpolation along a segment. Such interpolation could represent tapered or piecewise-varying primitives, but it would require new integration and compositing analyses. The work also leaves open whether feedforward predictors can infer these polylines robustly, and whether the recovered curves preserve enough physical fidelity to support inverse estimation of material or mechanical properties.

Conclusion

Spaghetti Splatting extends differentiable splatting from point supports to convolved segments and connected polylines. Its main technical contribution is a closed-form Gaussian ray integral for segment primitives, combined with an efficient EWA-based rasterization approximation and Mip-Splatting-style anti-aliasing. The zero-length limit recovers point splatting, while finite-length segments provide explicit continuity and improved representation of elongated geometry.

Across general scenes, the method matches or exceeds the evaluated point-based baselines while retaining comparable throughput. On hair reconstruction, it improves both photometric metrics and strand-level geometric F1, with a 2.7 dB PSNR advantage and a 14.3-point improvement at 2 mm. The paper therefore establishes convolutional curvilinear primitives as a practical extension of splatting, while leaving the reliability of topology recovery, high-frequency appearance modeling, and real-world strand reconstruction as specific unresolved questions.

Whiteboard

Explain it Like I'm 14

1. What is the paper about?

This paper introduces a new way to represent and render 3D objects called Spaghetti Splatting.

Many computer-graphics methods represent a 3D scene using tiny dots called points. This works well for many objects, but it is not ideal for things that naturally look like lines or strands, such as:

  • Hair
  • Fur
  • Grass
  • Wires
  • Cables
  • Thin branches
  • Brush bristles

The researchers replace isolated dots with connected line segments. These segments can join together to form long, bendy curves—much like strands of spaghetti.

The goal is to create realistic images of 3D scenes from new camera viewpoints while also recovering useful 3D curves.

2. What questions are the researchers asking?

The paper focuses on several main questions:

  1. Can curves be used instead of points to represent 3D scenes?
  2. Can these curves be trained directly from ordinary photographs?
  3. Can the method render images quickly enough for practical use?
  4. Does using connected curves improve the reconstruction of thin objects such as hair and fur?
  5. Can the same representation work for both thin strands and larger, ordinary objects?

The researchers also ask whether their curve-based method can remain compatible with successful point-based methods such as 3D Gaussian Splatting.

3. How does the method work?

From dots to soft shapes

A mathematical point has no width. If we tried to render a perfect point directly, it would be difficult to calculate how much of a pixel it should cover.

To solve this problem, graphics methods usually make each point slightly blurry or “soft.” Imagine placing a small cloud around every dot. The cloud tells the computer how much influence the dot has on nearby pixels.

This process is called convolution. In everyday language, convolution means taking a basic shape and spreading it out using a chosen pattern.

For example:

  • A point spread using a Gaussian kernel becomes a smooth, bell-shaped blob.
  • A point spread using a Beta kernel becomes another type of soft blob with a different shape.

Turning a line into a soft tube

Spaghetti Splatting applies the same idea to a line segment.

Instead of starting with one point, the method starts with a line between two 3D endpoints. It then spreads the line using a soft kernel. The result is similar to a thin, fuzzy tube.

Several connected segments form a polyline, which is a shape made from many straight pieces joined together. If the pieces are short enough, the polyline looks like a smooth curve.

Rendering the scene

To create an image, the method sends imaginary lines called camera rays from the camera into the scene. Each ray collects information from the objects it passes through.

The method calculates:

  • How much each curve contributes to the ray
  • How much light or color is blocked by earlier curves
  • Which curves are closer to the camera
  • The final color of each pixel

This process is called volume rendering. A simple analogy is looking through several partly transparent colored sheets. The front sheets affect what you see from the sheets behind them.

The researchers derive a shortcut, called a closed-form solution, for calculating the effect of a Gaussian-blurred line. This is important because the computer does not need to use many slow approximations.

Training the representation

The researchers begin with many curve segments and adjust them so that rendered images match real photographs.

This uses gradient descent, a training method similar to repeatedly making small corrections:

  1. Render an image.
  2. Compare it with the real image.
  3. Measure the difference.
  4. Move the curves and change their widths and colors to reduce the difference.
  5. Repeat many times.

The method can also split long segments, copy short segments, remove unhelpful ones, and join nearby segments into longer connected strands.

Reducing visual defects

The method includes anti-aliasing, which reduces jagged edges, flickering, and other problems that occur when very thin objects appear smaller than a pixel.

It also uses different kernels, including Gaussian and Beta kernels, to control the shape and spread of each curve.

4. What did the researchers find?

The experiments tested the method on general scenes, thin and fuzzy objects, and hair.

General scenes

On the Mip-NeRF 360 dataset, the Beta-kernel version of Spaghetti Splatting performed better than the comparison methods according to the paper’s image-quality measurements.

On the Shelly and Fuzzy datasets, the Gaussian-kernel version performed especially well.

The method was also fast. Its rendering speed was similar to that of point-based 3D Gaussian Splatting.

Thin and fuzzy objects

The method was able to represent objects such as:

  • Fur
  • Bristles
  • Spines
  • Fine leaves
  • Other narrow structures

A major advantage was that the method produced continuous curves rather than a cloud of separate points. This makes the reconstructed geometry look more like the real object.

Hair reconstruction

The researchers compared their method with Gaussian Haircut, a method designed specifically for reconstructing hair.

Spaghetti Splatting performed better in both image quality and strand accuracy. According to the paper, it achieved:

  • About 2.7 higher PSNR, a common image-quality score
  • A roughly 14-point improvement in F1 score at the strictest hair-accuracy setting

In simple terms, the reconstructed hair strands were closer to the true hair strands.

Connected structure

The experiments showed that joining segments together helped create long, organized strands.

Removing the joining process did not always greatly reduce image quality, but it produced many small disconnected pieces. The resulting geometry was less useful as a real strand structure.

A useful generalization of point splatting

A point can be thought of as a line segment whose length is zero. Therefore, Spaghetti Splatting includes ordinary point primitives as a special case.

This means the method can represent:

  • Ordinary point-like regions
  • Long thin curves
  • Mixtures of points and curves

That makes it more flexible than a method that can use only points or only lines.

5. Why are these results important?

The research is important because images alone usually do not directly tell a computer how objects are physically connected.

For example, a point-based method might show the appearance of a strand of hair using many scattered dots. Those dots may create a convincing image, but they do not necessarily form an actual strand that could be moved or simulated.

Spaghetti Splatting instead creates connected curves. This could make the recovered 3D objects useful for:

  • Animating hair and fur
  • Simulating wires, ropes, and flexible materials
  • Editing 3D scenes
  • Creating more realistic digital characters
  • Reconstructing natural thin structures
  • Building graphics tools for artists

Because the curves are connected, they can potentially be passed directly into physics simulations. For example, a simulated wind effect could move a reconstructed hair strand or cable.

6. Limitations and future improvements

The method is promising, but it is not perfect.

The paper identifies several limitations:

  • Fine textures may require many short curve segments.
  • Real hair is still difficult to reconstruct accurately.
  • The method may not always recover the true physical connections or topology of an object.
  • The experiments focus mainly on reconstruction from images, not on automatically predicting the curves with a single neural network.
  • More efficient ways to represent surface textures could reduce the number of required segments.

The researchers suggest combining the curves with texture maps, improving hair-specific training, and using the recovered geometry to estimate physical properties such as stiffness or flexibility.

Conclusion

Spaghetti Splatting is a method for representing 3D scenes with soft, connected curves instead of only separate points. It spreads each curve into a small volume, allowing the curves to be trained from photographs and rendered efficiently.

The experiments show that it can produce high-quality images, handle thin objects well, and reconstruct hair-like strands more accurately than some existing methods. Its biggest advantage is that it creates connected 3D structures that could later be edited, animated, or used in physics simulations.

In short, the paper moves computer graphics from thinking of thin objects as collections of scattered dots toward representing them as actual connected strands.

Practical Applications

Immediate Applications

  • Photorealistic 3D asset creation for film, games, and virtual production — graphics and entertainment. The method can reconstruct novel-view-renderable scenes directly from multiple photographs while representing elongated structures as connected polylines rather than disconnected points. Production teams could use it to capture hair, fur, grass, foliage, brush bristles, wires, cables, ropes, and other thin geometry for rapid asset generation. Its rendering speed, comparable to point-based splatting, supports interactive preview and camera exploration. Dependencies: sufficiently overlapping multi-view images, calibrated or estimable cameras, GPU-based rendering, and scene-specific optimization. Fine textures may still require many short segments or separate texture maps.
  • Digital hair and fur reconstruction for character pipelines — animation, games, and VFX. Reconstructed strands can be exported as connected curves and passed directly to hair-shading and position-based dynamics systems. A practical workflow is: capture multi-view images of a hairstyle or furry object, optimize the polyline representation, convert the resulting strands into a simulation-ready hair format, and refine roots, density, and grooming artistically. The reported improvement over Gaussian Haircut suggests particular value for curly and highly curved hair. Dependencies: accurate strand-root estimation, reliable segmentation of hair from the head or object, adequate view coverage, and additional treatment of occlusions, lighting, and physically plausible hair density.
  • Interactive visualization of thin and fibrous objects — CAD, inspection, and digital twins. The representation can provide fast renderable models of cables, tubing, fibers, needles, plant stems, bristles, and similar structures. Connected primitives preserve strand continuity better than scattered point clouds and can remain stable across viewing distances because of the integrated anti-aliasing process. Dependencies: the current method is optimized primarily for appearance reconstruction rather than guaranteed metric accuracy; engineering use requires calibration, scale validation, and uncertainty estimates.
  • Simulation-ready reconstruction for robotics and physical computing — robotics and embodied AI. Since the output consists of long connected strands, it can be inserted into physics engines for collision, deformation, and position-based simulation. Potential uses include reconstructing ropes, cables, soft tubing, cloth-like fibers, vegetation, or hair around robots and humans. A robot could use the reconstructed geometry for collision-aware manipulation or scene interaction. Dependencies: the paper demonstrates amenability to simulation but does not establish physical parameter recovery. Mass, stiffness, friction, attachment points, and collision tolerances must be estimated separately.
  • Fast multi-view capture and asset scanning — 3D content production. The closed-form ray integration for Gaussian-convolved segments and the segment-wise EWA approximation can reduce the cost of training and inference relative to numerical volume integration. Capture studios could use the released implementation as a reconstruction backend for rapid inspection of scanned objects before producing higher-quality meshes or strand models. Dependencies: performance depends on GPU hardware, primitive budgets, scene complexity, and implementation quality. The method remains a per-scene optimization approach rather than a universal feed-forward scanner.
  • Anti-aliased rendering of distant thin structures — visualization and remote inspection. The integrated 3D and 2D filtering is suitable for rendering structures that project to subpixel widths, such as wires, antennae, grass, bristles, or hair. This can reduce flickering and aliasing in animation, interactive 3D viewers, and camera fly-throughs. Dependencies: the anti-aliasing behavior is tied to the sampling assumptions of the training views and to the screen-space linearization used by the renderer. Extreme perspective changes or highly detailed textures may require additional filtering.
  • A research and teaching platform for differentiable rendering — academia. The method provides an accessible experimental framework for studying how convolution kernels and geometric supports affect differentiable rendering. Researchers can vary the support—points, segments, and polylines—and the kernel—Gaussian, Beta, or other profiles—within a common formulation. The public code can support coursework and experiments in computer vision, graphics, inverse rendering, and numerical optimization. Dependencies: the paper text contains apparent anonymization and LaTeX corruption, so users should rely on the final released implementation and documentation rather than the manuscript transcription alone.
  • Reusable rendering components for software libraries — graphics software. The method can be integrated into differentiable renderers, neural scene-representation libraries, and GPU splatting frameworks as a polyline primitive with closed-form opacity evaluation. Its strict generalization of point primitives allows software to support both zero-length points and connected segments in one representation. Dependencies: production integration requires robust handling of sorting, memory management, dynamic densification, topology changes, gradients at segment joins, and interoperability with standard mesh and curve formats.
  • Policy and cultural-heritage visualization — public-sector imaging and documentation. Agencies and museums could use multi-view captures to create viewable digital records of fine structures such as woven artifacts, textile fibers, plant specimens, or architectural cables. The representation is particularly useful when preserving thin geometry is more important than creating a conventional watertight mesh. Dependencies: archival use requires provenance metadata, reproducible capture conditions, scale references, long-term software compatibility, and explicit separation between visually plausible reconstruction and verified physical structure.
  • Consumer-facing 3D capture for virtual meetings and commerce — daily life and consumer software. A mobile or desktop application could reconstruct a person, pet, hairstyle, houseplant, or object from a short multi-view capture and provide interactive novel views. The connected-strand representation may improve the appearance of hair, fur, and plant-like objects compared with point-only methods. Dependencies: current training is not demonstrated in real time or on mobile devices; camera calibration, privacy-preserving processing, computational cost, and difficult lighting conditions remain practical barriers.

Long-Term Applications

  • Feed-forward reconstruction of curves from a small number of images — computer vision and edge AI. The paper identifies feed-forward prediction as an important future direction. A learned model could directly infer polyline endpoints, connectivity, kernel widths, densities, colors, and appearance from one or a few images, replacing per-scene gradient-based optimization. This could enable near-real-time reconstruction for mobile scanning, robotics, and augmented reality. Dependencies: large and diverse training datasets with strand-level or curve-level ground truth, robust handling of occlusion and ambiguous topology, and methods for estimating uncertainty are required.
  • Physics-aware digital twins of deformable objects — robotics, manufacturing, and simulation. Reconstructed curves could be combined with learned physical properties to create digital twins of cables, hoses, fibers, ropes, hair, and plant stems. The system could infer not only geometry but also stiffness, damping, friction, and attachment constraints, enabling simulation-based planning and predictive maintenance. Dependencies: the current method recovers appearance-oriented geometry, not material parameters. Ground-truth force measurements, temporal observations, differentiable physics, and domain-specific calibration would be needed.
  • Robotic manipulation of ropes, wires, and flexible materials — robotics. A robot could use the connected geometry to estimate the configuration of a cable or rope, plan grasp points, detect entanglements, and predict future configurations. The continuity of polylines is potentially more useful for manipulation than an unordered collection of Gaussian points. Dependencies: real-time inference, temporal tracking, robust occlusion handling, metric accuracy, and integration with tactile or depth sensors are necessary. Safety-critical deployment would require formal error bounds.
  • Curve-to-surface and hybrid geometric modeling — CAD, manufacturing, and scientific visualization. The proposed kernel generalization could be extended from uniform segment kernels to kernels that vary along a segment. This may produce cones, tapered fibers, triangles, trapezoids, and eventually surface-like primitives. A hybrid renderer could combine curves for thin structures, triangles for surfaces, and points for unresolved detail. Dependencies: endpoint-interpolated kernels, stable gradients, topology management, and validation against explicit meshes remain to be developed.
  • High-quality digital humans and virtual try-on — healthcare, retail, and entertainment. Improved strand reconstruction could support realistic hair in digital humans, avatar generation, virtual try-on, telepresence, and surgical or cosmetic planning. Combined with learned priors, the system could infer plausible hidden hair regions and maintain consistent strand flow under changes in viewpoint or pose. Dependencies: privacy protection, identity and biometric safeguards, temporal consistency, accurate head and scalp geometry, learned priors for occluded regions, and validation across hair types and lighting conditions.
  • Plant and ecological monitoring — agriculture, forestry, and environmental science. The approach could reconstruct connected stems, tendrils, grasses, roots, and fine branches from image sequences. Such models could support growth tracking, canopy analysis, biomass estimation, and structural phenotyping. Dependencies: biological structures may be occluded, deform over time, and vary in scale. Scientific use requires calibrated measurements, temporal reconstruction, species-specific validation, and uncertainty quantification rather than image-quality metrics alone.
  • Medical and biomedical modeling of tubular or filamentary anatomy — healthcare. In principle, connected differentiable curves could represent vessels, airways, nerves, catheters, or surgical sutures in volumetric visualization and simulation. The differentiable rendering formulation could help fit these structures to multi-view or multimodal imaging data. Dependencies: the paper does not evaluate medical data, and clinical deployment would require high geometric accuracy, modality-specific imaging models, patient privacy, regulatory validation, and safeguards against hallucinated anatomy. This is therefore a research direction, not an immediate clinical application.
  • Inverse rendering and material-property estimation — graphics, industrial inspection, and science. Once the geometry is explicitly represented as connected curves, optimization could be extended to infer width, density, reflectance, anisotropy, and other physical properties from images. This could support inspection of fiber composites, textiles, coatings, vegetation, and manufactured cables. Dependencies: appearance and geometry are coupled in the current formulation; additional illumination variation, polarization or depth data, material priors, and identifiability analysis would be needed to distinguish shape from reflectance.
  • Temporal and deformable Spaghetti Splatting — video understanding and AR/VR. Extending the method to time-dependent polylines could produce stable strand identities and track deformation through video. Applications include animated hair, moving ropes, cloth fringes, dynamic plants, and flexible robot components. Dependencies: temporal regularization, correspondence across frames, topology changes, motion blur, and efficient incremental optimization must be addressed. Independent per-frame reconstruction would likely cause strand flicker and inconsistent connectivity.
  • Standardized interoperable curve assets for production pipelines — software and digital content standards. A mature implementation could export reconstructed strands to USD, Alembic, Blender, game engines, and physics packages, preserving connectivity, radius, density, color, and uncertainty. This would make differentiable reconstruction a front end for conventional production and simulation workflows. Dependencies: consensus on representation semantics, stable conversion to renderer-specific curve formats, licensing of datasets and code, and preservation of the distinction between rendering primitives and physically verified geometry.
  • Policy tools for infrastructure inspection — energy, transportation, and public works. Future systems could reconstruct power lines, cables, bridge hangers, pipelines, vegetation encroachment, and other elongated infrastructure from drone or vehicle imagery. Differentiable fitting could support change detection and prioritization of maintenance. Dependencies: metric scale, georeferencing, regulatory-grade accuracy, adverse-weather robustness, and independent sensor validation are essential. The paper’s image-quality benchmarks do not by themselves establish inspection reliability.
  • Learning structured priors for topology and connectivity — academia and scientific machine learning. The paper explicitly cautions that automatically recovered topology has not been broadly validated. Future research could develop priors that enforce rooted strands, branching structures, loop constraints, or known connectivity, enabling more reliable reconstruction of hair, plants, cables, and biological filaments. Dependencies: topology-aware losses, annotated datasets, evaluation beyond PSNR/SSIM/LPIPS, and methods for representing uncertainty over alternative connectivities are required.

Glossary

  • Anti-aliasing: A rendering technique that reduces visual artifacts caused by insufficient sampling of high-frequency image details. “To avoid aliasing, we follow Mip-Splatting~\cite{yu2024mipsplatting} and apply a 3D as well as a 2D Gaussian filter to our primitives.”
  • Anisotropic kernel: A kernel whose spread varies according to direction. “The first column shows that convolving a point mass with a Gaussian or a Beta kernel yields the primitives of 3DGS~\cite{kerbl2023gaussians} and DBS~\cite{liu2025dbs}, respectively.”
  • Back-propagate: To compute gradients through a computational process so that parameters can be optimized. “The segments are then composited along each ray to obtain the accumulated color.”
  • Beta kernel: A compact kernel based on the beta probability distribution, used to model localized primitive influence. “The formulation in \cref{eq:primitive} extends naturally to other convolution kernels, such as the beta kernels of~\cite{liu2025dbs}.”
  • Closed-form integral: An integral that can be evaluated exactly using a finite mathematical expression rather than numerical approximation. “We now describe how our primitive defines a volumetric density field, and, without loss of generality, derive the closed-form volume rendering integral for the Gaussian kernel.”
  • Composite: To combine multiple partially transparent rendered contributions in depth order. “For polylines, we treat each segment independently: we evaluate the volume integral and the resulting opacity separately for every segment, and then composite the segments along the ray as individual primitives.”
  • Convolution: An operation that combines a signal or measure with a kernel to produce a spatially distributed representation. “We make curves differentiable by convolving them with a kernel, and show that the convolution of a polyline converges to that of the continuous curve as its segments shrink.”
  • Convolution surface: An implicit surface obtained from the convolution of a geometric skeleton with a smooth kernel. “A classical computer graphics technique, Convolution Surfaces~\cite{bloomenthal1991convolution}, has shown that one can represent many implicit shapes as an iso-surface of a convolution of a skeleton of curves and polygons with a smooth kernel.”
  • Covariance: A matrix describing the scale and orientation of a multivariate distribution or kernel. “Here, x∈R3x{\in}\mathbb R^3 is a query point, δ\delta is the three-dimensional Dirac distribution, isthecovarianceofthekernel,andis the covariance of the kernel, and denotes convolution.”
  • Curvilinear primitive: A geometric rendering primitive that follows a curved path, such as a polyline or curve. “As points~\cite{kerbl2023gaussians,yu2024mipsplatting,liu2025dbs} and triangles~\cite{held2026triangles,held2025meshsplatting} have already received significant attention, we focus on differentiating through curvilinear primitives; see \Cref{fig:convolution}.”
  • Densification: The adaptive addition, splitting, cloning, or relocation of primitives to increase representational capacity. “As in other per-scene optimization methods, we adaptively resample our primitives to allocate representation capacity where it is needed.”
  • Dirac distribution: A generalized function concentrated at a single point, commonly used to represent an idealized point mass. “Here, x∈R3x{\in}\mathbb R^3 is a query point, δ\delta is the three-dimensional Dirac distribution, isthecovarianceofthekernel,andis the covariance of the kernel, and denotes convolution.”
  • EWA splatting: Elliptical weighted average splatting, a filtering and rasterization method that projects spatial kernels onto the image plane. “EWA splatting~\cite{zwicker2002ewa} and [PERSON_NAME] reconstruction~\cite{kazhdan2006poisson} both turn a [PERSON_NAME]-like representation into a volumetric one by convolving it with a low-frequency kernel.”
  • F1 score: The harmonic mean of precision and recall, used to evaluate prediction or reconstruction accuracy. “We report the F1 score based on this criterion at the thresholds 2\,mm and 20∘20^\circ, 3\,mm and 30∘30^\circ, and 4\,mm and 40∘40^\circ~\cite{sklyarova2023neuralhaircut}.”
  • Foreshortening: The apparent shortening of an object caused by its orientation relative to the viewing direction. “The yellow segment in \Cref{fig:method} shows the resulting foreshortening.”
  • Front-to-back compositing: Rendering in which contributions are accumulated in order from the closest surface or interval to the farthest. “Splitting the ray into ordered intervals IkI_k from front to back, with a constant color ck\bm c_k on each, turns \cref{eq:vre} into front-to-back compositing.”
  • Gradient descent: An optimization method that iteratively updates parameters in the direction that decreases an objective function. “Rather than constructing them by hand, we would like to optimize such primitives directly from a collection of images, by volume rendering~\cite{mildenhall2020nerf} and gradient descent.”
  • Implicit surface: A surface represented as a level set of a scalar field rather than explicitly by vertices or polygons. “In both cases, the convolved field serves only to define an iso-surface.”
  • Iso-surface: The set of points at which a scalar field has a specified constant value. “A classical computer graphics technique, Convolution Surfaces~\cite{bloomenthal1991convolution}, has shown that one can represent many implicit shapes as an iso-surface of a convolution of a skeleton of curves and polygons with a smooth kernel.”
  • Kernel: A function used in convolution to distribute or weight the influence of a geometric primitive. “We make curves differentiable by convolving them with a kernel, and show that the convolution of a polyline converges to that of the continuous curve as its segments shrink.”
  • Low-frequency kernel: A smoothing kernel that suppresses rapid spatial variations and high-frequency details. “EWA splatting~\cite{zwicker2002ewa} and [PERSON_NAME] reconstruction~\cite{kazhdan2006poisson} both turn a [PERSON_NAME]-like representation into a volumetric one by convolving it with a low-frequency kernel.”
  • Marginal: A distribution obtained by integrating out one or more variables from a joint distribution. “where Q=12−1Q{=}\tfrac12^{-1} and $Q_\perp{=}Q-\mQdd^\topQ/(d^\top\mQd)$ is the matrix QQ of the marginal of NN_{} along the camera ray.”
  • Multisample anti-aliasing: An anti-aliasing method that evaluates multiple samples within each pixel. “In contrast, the concurrent work of \citet{tojo2026lines} does not support spatial convolutions and must therefore resort to multisample anti-aliasing with four samples per pixel.”
  • Nyquist–[ADDRESS] theorem: A sampling theorem stating that a signal must be sampled at more than twice its highest frequency to be reconstructed without aliasing. “This makes sampling the representation onto a discrete grid of pixels well posed from the perspective of the Nyquist--[ADDRESS] theorem~\cite{yu2024mipsplatting}.”
  • Opacity: The fraction of light blocked or absorbed by a medium or rendered primitive. “We now turn from the density to the opacity that a primitive contributes along a camera ray.”
  • Optical depth: The accumulated density along a light or camera ray, determining attenuation or transmittance. “Substituting \cref{eq:f-gaussian} into \cref{eq:tau-f} reduces the optical depth to a single integral.”
  • Polychromatic: Relating to multiple wavelengths or colors of light; in this paper, the term is relevant to color-field rendering. “Given a density field σ(⋅)\sigma(\cdot) and an emitter color field c(⋅)\bm c(\cdot), the color along a ray r(t)=o+td\bm r(t)=o+td as per volume rendering~\cite{max1995optical} can be computed as:”
  • Polyline: A connected sequence of line segments used to approximate a continuous curve. “We approximate curves by polylines, and volume render each segment so that connected segments composite correctly.”
  • Primitives: Basic geometric elements, such as points, lines, triangles, or segments, used to represent a scene. “In production graphics systems, 3D geometry is represented by points, lines, and triangles~(\ie, primitives), as rendering hardware is optimized to generate images from them very efficiently.”
  • Radiance field: A spatial function that describes the emitted or observed color and density of a scene from different positions and directions. “We introduce a differentiable 3D representation which can be used to learn accurate radiance fields for novel view synthesis applications~(left).”
  • Rasterization: The process of converting geometric or continuous scene descriptions into a discrete grid of pixels. “Their method rasterizes segments on a subpixel grid using the [PERSON_NAME] algorithm and recovers high-quality geometry in object-centric scenes.”
  • Ray integral: An integral that accumulates a function along a line corresponding to a camera ray. “The ray integral of a Gaussian-convolved segment only involves the [ADDRESS] error function~erf⁡\operatorname{erf}~(\ie, the anti-derivative of a Gaussian).”
  • Spherical harmonics: A set of basis functions defined on the sphere, commonly used to represent directional signals. “Spherical Voronoi replaces spherical harmonics with a learned directional partition~\cite{disario2026sv}.”
  • Splatting: A rendering technique that projects and accumulates the influence of primitives or kernels onto an image plane. “It is splatting, as each primitive is rendered by splatting its kernel onto the image.”
  • Strict superset: A set or representation that contains another set while also including additional elements not present in it. “As a point is a curve of zero length, our representation is a strict superset of 3D [PERSON_NAME].”
  • Subpixel: Smaller than a single image pixel. “Their method rasterizes segments on a subpixel grid using the [PERSON_NAME] algorithm and recovers high-quality geometry in object-centric scenes.”
  • Transmittance: The fraction of light that remains after passing through an absorbing or attenuating medium. “where the transmittance term is defined as”
  • Volume rendering: The process of generating images by integrating color and density through a volumetric scene representation. “Neural Radiance Fields~(NeRF) represents a scene as a continuous density and color field rendered by numerical integration along camera rays~\cite{mildenhall2020nerf}.”
  • Volumetric density field: A spatially varying scalar field that specifies the density of matter or opacity throughout a three-dimensional volume. “We now describe how our primitive defines a volumetric density field, and, without loss of generality, derive the closed-form volume rendering integral for the Gaussian kernel.”
  • Zero-length limit: The limiting case in which a geometric segment shrinks to a single point. “We introduced a splatting representation formed by convolving finite segments with anisotropic kernels, with anisotropic point primitives recovered in the zero-length limit.”

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Tweets

Sign up for free to view the 2 tweets with 418 likes about this paper.