NURBS Splatting: A Unified Differentiable Rendering Framework for Vector Graphics

arXiv:2606.31764 · cs.GR, cs.CV · Submitted 2026-06-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "NURBS Splatting: A Unified Differentiable Rendering Framework for Vector Graphics".

Tom: Differentiable rendering of planar rational splines remains largely underexplored, despite their widespread use in vector graphics and design.

Jane: First, who's behind it and why it matters.

Paper summary: Tom: So, wrapping up this discussion on NURBS Splatting: the authors are pushing a unified framework that lets you render planar rational splines by treating them as continuous Gaussian fields. It’s a big step because it directly addresses the limitations of previous differentiable renderers that were stuck with polynomial Bézier curves.

Jane: It seems like this paper is important because it provides a pathway for high-quality vector graphics and image abstraction by supporting features like rational weights and non-uniform knots. It moves the technology closer to how CAD systems actually model things.

Lu: The implication here is that we can finally get a smooth, differentiable representation of shapes that are natively defined in industry tools, which opens up new possibilities for integrating AI generation with professional design workflows.

Meng: Practically speaking, this means if we can use this framework to vectorize images layer by layer or abstract them using diffusion models, the resulting vector data will be much more accurate and usable.

Lalam: I see huge potential here for our AI culture; imagine generating intricate designs where the underlying geometry is mathematically perfect, not just an approximation of a Bézier path. That level of structural fidelity in generated art will be something we can really leverage.

Tom: Absolutely; the authors demonstrate this effectiveness across calligraphy reconstruction and image vectorization frameworks, showing improvements over existing approaches. It’s about taking a technique that was already powerful, Gaussian splatting, and applying it to a new geometric primitive.

Jane: And they show speed improvements too; for instance, coupling it with diffusion-based Score Distillation Sampling achieves stylization quality comparable to prior methods while running about one point two seven times faster.

Lu: The framework’s ability to support per-control-point shape modulation for tighter local control, thanks to those rational weights, is what makes the geometric precision achievable here.

Meng: My main concern is always scalability; how well this framework handles very large datasets or extremely complex, long splines in a real-world application environment.

Tom: That’s a fair point, Meng; the authors did mention that the system exhibits improved scalability compared to methods that require pre-conversion into Bézier curves. It suggests they've thought about efficiency during the design phase.

Jane: Overall, NURBS Splatting seems to be a solid development because it bridges the gap between high-fidelity CAD modeling and modern differentiable rendering techniques.

Lalam: It’s exciting because this means the tools we build won't just be about making things look good; they will be about making them geometrically sound and perfectly represented in a way that respects the underlying mathematical structure of design.

Conclusion: Tom: So, we've been diving deep into how NURBS Splatting works and what it achieves for vector graphics rendering. Now we're wrapping up with Tom and Jane discussing the title and authors of this paper and what all this means for us out there in the real world.

Jane: It sounds like a lot to take in, Tom; I'm trying to make sure I grasp how these complex ideas translate into something practical for our listeners.

Lu: From my perspective, the title itself really captures the essence: it’s about unifying differentiable rendering for planar rational splines. That unification is what makes this approach so interesting for AI applications in design.

Meng: Unification is a big word, Lu; from an engineering standpoint, I'm more interested in how they managed to tie together Gaussian fields with control points and weights so seamlessly within a single differentiable pipeline.

Lalam: The authors are really tackling the problem of making these mathematically rich shapes usable by modern machine learning models for both creation and understanding. It feels like a major step toward creating design tools that are inherently intelligent rather than just static representations.

Tom: Exactly, Lalam; it’s not just about rendering anymore; it’s about building systems where the geometry itself is learnable and optimizable through image loss. The authors have clearly put a lot of thought into making this framework robust enough to handle those complex curves we use every day.

Jane: I think what's most important for our listeners to understand is that this method allows us to optimize the shape, the weights, and even the knot vectors simultaneously based on how well it reconstructs an image. It’s a self-correcting system, essentially.

Lu: That self-correction capability is what excites me; we can move beyond simple path tracing or texture mapping and into a realm where AI can directly sculpt high-fidelity vector data from abstract visual concepts.

Meng: But I still wonder about the practical side: how easy is it for someone on the ground to take this framework and deploy it in a production environment without needing an army of researchers to tune all those loss functions?

Lalam: That's where the cultural impact really hits home; if we can simplify this level of geometric control, it could democratize high-quality generative design, letting more creators access tools that understand complex shapes intuitively.

Tom: Well said, Lalam; it’s about moving the complexity from the end-user to a sophisticated optimization loop running on the machine. This paper shows we can handle those complexities without losing control over the final output's fidelity.

Jane: So, in simple terms, these authors have created a unified way to render intricate vector shapes by letting an image reconstruction loss guide every single parameter of the curve itself.

Lu: Precisely; they’ve turned a complex geometric problem into a smooth optimization problem solvable with standard differentiable rendering tools. That's the core mechanism here.

Meng: It’s impressive that they managed to bake in derivative losses to enforce regularity on the curves while still allowing for that massive optimization space of weights and control points. I'm curious if those regularization terms add significant computational overhead in practice, though.

Lalam: The ability to use non-uniform knots and rational weights means we can achieve a level of precision in shape representation that was previously only possible with painstaking manual modeling, which is huge for cultural artifacts too.

Tom: And the results speak for themselves; they showed tangible improvements in calligraphy reconstruction and image vectorization over existing methods, proving this isn't just theoretical fluff.

Jane: So, while the paper focuses on a unified framework, its implications are that we can expect vector graphics and AI-driven content creation to become significantly more accurate and geometrically expressive very soon.

Lu: The future work I see involves pushing this further into even more complex topological spaces; imagine applying this to three dee surfaces defined by rational splines in the next phase.

Meng: That sounds ambitious, Lu; for now, I'm focused on seeing how quickly we can prototype a deployable version of the core pipeline we discussed.

Lalam: And from where I sit, this work suggests that the next big cultural shift won't just be about generating pretty pictures, but about generating perfectly structured digital objects that respect mathematical reality.

Tom: That's a powerful vision to leave us with; it really puts the scale of what we're talking about into perspective.

Jingye Qiu, Shizhe Zhou

Hunan University

cs.GR, cs.CV

Submitted: 2026-06-30

Updated: 2026-09-29

Code: https://github.com/AnicoderAndy/nurbs-splatting

Importance score: 79/100

The gist: Differentiable rendering of planar rational splines remains largely underexplored, despite their widespread use in vector graphics and design.

Key concepts

Gaussian Splatting
This technique reformulates rendering by sampling isotropic Gaussians along or within regions defined by NURBS curves. This process produces a differentiable image, meaning changes in the image directly inform how to adjust the underlying geometric parameters.
Rational Splines
These are complex vector graphics curves that use rational weights and non-uniform knot vectors. Unlike simpler polynomial Bézier curves, NURBS splines can exactly represent conic sections and offer per-control-point shape modulation for finer local control.
Differentiable Rendering Pipeline
The entire rendering process is made differentiable so that image reconstruction losses drive gradients back to the geometric inputs, such as control points and weights. This allows the system to learn the optimal curve structure directly from visual data.
Geometric Regularization Loss
This loss term penalizes high-frequency variations in the underlying curves by measuring their derivatives. By enforcing smoothing constraints on these derivatives, it ensures that the optimized splines maintain a structurally sound and regular shape.

Terminology

Summary

Differentiable rendering of planar rational splines remains largely underexplored, despite their widespread use in vector graphics and design.

The gist

NURBS Splatting proposes a unified framework that represents planar rational curves as continuous Gaussian fields, allowing for the differentiable rendering of these complex geometric primitives in 2D image space while enabling joint optimization of control points, rational weights, and non-uniform knot vectors.

How it works

The framework reformulates rendering by sampling isotropic Gaussians along NURBS curves or within closed regions using an SDF-modulated grid. This process is managed through a tile-based Gaussian splatting rasterizer to produce a differentiable image. The pipeline is fully differentiable, meaning image-space losses drive gradients back to control points, rational weights, knot intervals, stroke widths, colors, and opacities.

The method handles different curve types by employing distinct sampling strategies:

  1. For open curves: Gaussians are placed adaptively along the contour with density proportional to arc length. The Gaussian center is computed via a span-local formulation exploiting the compact support of B-spline basis functions, reducing evaluation cost from O(Mn) to O(Mp).

  2. For filled closed regions: Interior primitives are sampled using a uniform 2D grid within an axis-aligned bounding box, with opacity modulated by a differentiable signed distance field (SDF). This involves sampling boundary points and computing the SDF based on the winding number.

Optimization and Regularization

The framework optimizes a combined objective function consisting of an image-space reconstruction loss and geometric regularization:

  1. Image Loss: A loss, such as MSE, drives gradients through the Gaussian parameters back to the NURBS control points and weights.

  2. Derivative Loss: This term enforces regularity on the underlying curves by penalizing the squared magnitude of the r-th derivative of a parametric curve averaged over its domain, specifically using the third-order derivative smoothing loss for degree 3 curves.

  3. Bounding Box Loss: This penalty prevents control points from drifting outside the target region defined by the image's bounding box.

Applications and Results

The framework has been demonstrated across several vector graphics tasks, showing improvements over polynomial baselines:

  1. Calligraphy Reconstruction: The method recovers both stroke trajectories and varying widths by optimizing control point positions, rational weights, knot intervals, and per-control-point widths jointly. It shows a 33% reduction in MSE compared to the baseline.

  2. Layer-wise Image Vectorization (LIVE): By replacing cubic Bézier paths with NURBS curves and introducing a grid-step annealing strategy, the method achieves higher quality and speed than the LIVE baseline, demonstrating robustness to complex topologies.

  3. Neural Image Abstraction: The framework is coupled with a diffusion-based Score Distillation Sampling (SDS) pipeline, achieving comparable stylization quality to prior methods while running 1.27× faster.

Key Advantages

The primary advantage of NURBS Splatting is its ability to support rational weights and non-uniform knots, which allows for the exact representation of conic sections and provides per-control-point shape modulation for tighter local geometric control, a capability missing in existing differentiable renderers restricted to polynomial Bézier curves. Furthermore, the use of isotropic Gaussian sampling eliminates artifacts like blurred edges at large stroke widths and spikes at high-curvature regions observed in anisotropic methods. The framework also exhibits improved scalability with respect to the number of control points compared to methods that require pre-conversion into Bézier curves.

Summary of Parameters

The system allows for the learnable optimization of:

(See Table S1 for full details)

Control points (Pi), rational weights (wi), and knot intervals (τj). The curve also carries an RGB color and opacity. For filled regions, the interior grid step size 'h' is annealed during optimization to accelerate convergence while preserving detail. Loss functions include MSE, the third-order derivative loss, and bounding box penalties. The framework provides a clean structure allowing manual adjustment of optimized control points for editability.

References

(See Section 15 for full list)

**(Note: The summary is structured to meet the requested length and format constraints, focusing only on information explicitly present in the text.

Improvements for AI systems

Based on the provided scientific paper, here are specific improvements to existing AI systems that could be achieved by implementing or integrating NURBS Splatting:

)A. Enhanced Vector Graphics Generation and Editing Systems (Calligraphy & Typography):

  1. Improved Calligraphy Reconstruction and Style Transfer:

The system can move beyond generating simple stroke outlines by accurately reconstructing complex, non-uniform calligraphy strokes with varying widths and high-curvature details. This is achieved by jointly optimizing control points, rational weights, and knot vectors.

  1. Accurate Geometric Fidelity for Vector Output:

Unlike polynomial B-spline methods which approximate conic sections poorly (e.g., circles), NURBS Splatting can render exact circular arcs and ellipses with high fidelity, crucial for professional font design and digital archiving where geometric accuracy is paramount.

  1. Differentiable Style Modulation:

The framework allows for the integration of style-guided losses (e.g., patch-wise directional CLIP loss) directly into the optimization loop, enabling AI systems to generate vector art that adheres strictly to a desired artistic style while maintaining geometric correctness, leading to higher quality and more controllable outputs than current methods.

(Specific Improvement: Replace existing Bézier/polynomial curve optimizers with NURBS Splatting for stroke extraction and generation.)

)B. Robust Image-to-Vectorization Pipelines (Live Vectorization):

  1. High-Fidelity Layered Vectorization:

The system can perform layer-wise image vectorization that produces significantly more robust and geometrically accurate closed paths, especially for complex topologies like facial expressions or thin interleaving structures, which often fail with current DiffVG baselines.

  1. Adaptive Optimization for Complex Scenes:

By incorporating grid-step annealing during the optimization of filled regions, the system can handle large-scale geometric deformations (like those in complex emojis) more effectively without sacrificing fine detail during early stages of training, leading to more reliable output across diverse image inputs.

(Specific Improvement: Upgrade image vectorization models from Bézier path generation to a NURBS Splatting rasterizer pipeline.)

)C. Generative Modeling and Neural Image Abstraction (Text-to-Vector/Image Synthesis):

  1. Semantic Vector Synthesis:

The system can generate high-quality, editable vector designs directly from text or image prompts by leveraging the differentiable nature of the NURBS representation. The ability to optimize rational weights allows for semantic control over local geometry, enabling AI to synthesize stylized vector art that matches complex textual descriptions with geometric precision (e.g., a black and white ink drawing).

  1. Accelerated Diffusion-Guided Abstraction:

When coupled with diffusion models (like ControlNet/IPAdapter), the system can achieve faster optimization of the underlying geometric parameters compared to purely polynomial baselines, allowing for faster synthesis of complex vector abstractions from raster inputs while maintaining high reconstruction quality.

(Specific Improvement: Integrate NURBS Splatting as the geometric primitive renderer within a diffusion model pipeline for text-to-SVG generation.)

)D. Computational Efficiency and Scalability:

  1. Improved Training Scalability with Complexity:

The system exhibits superior scalability concerning the number of control points compared to methods that rely on pre-processing conversions (like Bézier extraction). This means the AI can handle exponentially more complex vector structures (longer splines, higher detail) without incurring prohibitive overhead during the optimization phase.

(Specific Improvement: Utilize NURBS Splatting's direct NURBS representation instead of polynomial conversion steps to handle increasingly detailed geometric inputs.)

Sources

Related papers