---

VectorArk: Learning Practical Image Vectorization with Rounded Polygon Representation

SVGVectorizationVLMCVPR 2026Adobe

Gehlaut T, Liu D, Bansal C, et al. VectorArk: learning practical image vectorization with rounded polygon representation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2026: 31619-31627.

https://arxiv.org/abs/2605.24398v1

https://vectorark.github.io/

---

VectorArk：利用圆角多边形表示学习实用的图像矢量化

图1：我们的框架能够处理真实世界矢量化场景，包括文生图模型产生的退化输出，以及低分辨率的矢量图渲染结果。通过专门的圆角多边形表示与基于轮廓的流程，将几何预测与样式处理解耦；我们的方法生成干净、平滑的矢量图元，再在后处理步骤中为其赋予样式。

图1：我们的框架能够处理真实世界矢量化场景，包括文生图模型产生的退化输出，以及低分辨率的矢量图渲染结果。通过专门的圆角多边形表示与基于轮廓的流程，将几何预测与样式处理解耦；我们的方法生成干净、平滑的矢量图元，再在后处理步骤中为其赋予样式。

Abstract

Recent vision-language model (VLM)–based approaches have achieved impressive results on image vectorization tasks. However, they are typically evaluated on synthetic benchmarks, where clean SVGs are rasterized at high resolution and then re-vectorized. As a result, these methods generalize poorly to real-world scenarios, such as images with unknown rasterization methods or those generated by text-to-image models. We introduce VectorArk, a new VLM-based model designed for robust and practical image vectorization. VectorArk employs a novel rounded polygon representation that simplifies the learning process while naturally producing smooth, visually appealing primitives. We also propose a degradation model that enhances robustness across diverse and imperfect inputs. Our experiments show that, in contrast to previous methods, VectorArk achieves superior geometric completeness and artifact suppression across multiple datasets, with comprehensive ablations validating the contribution of each component.

近期基于视觉语言模型（VLM）的方法在图像矢量化任务上取得了令人瞩目的结果。 然而，这些方法通常在合成基准上评估：先将干净的 SVG 栅格化为高分辨率图像，再重新进行矢量化。 因此，这些方法在真实世界场景中的泛化能力较差，例如栅格化方式未知的图像，或由文生图模型生成的图像。 我们提出 VectorArk，一种基于 VLM 的新模型，旨在实现鲁棒且实用的图像矢量化。 VectorArk 采用一种新颖的圆角多边形表示，既简化了学习过程，又能自然生成平滑且美观的图元。 我们还提出一种退化模型，以增强对多样且不完美输入的鲁棒性。 实验表明，与以往方法相比，VectorArk 在多个数据集上实现了更好的几何完整性与伪影抑制效果，全面的消融实验也验证了各个组件的贡献。

1. Introduction

Converting raster images to vector graphics is a classical but challenging task in computer graphics. Compared with raster images, vector-based representations offer resolution-independent rendering, compact storage, and improved editability, making them widely used in applications such as vector graphics editors, web rendering engines, and data visualization tools. Classic image-to-vector approaches excel in reliability and computational efficiency, yet they frequently produce imprecise geometry or redundant paths, often lacking semantic awareness of the underlying layer structure.

将栅格图像转换为矢量图形，是计算机图形学中的经典问题，但仍具有挑战性。 与栅格图像相比，矢量表示支持与分辨率无关的渲染、紧凑存储以及更好的可编辑性，因此广泛用于矢量图形编辑器、网页渲染引擎和数据可视化工具等应用。 传统的图像转矢量方法在可靠性和计算效率方面表现出色，但经常生成不精确的几何形状或冗余路径，也往往缺乏对底层图层结构的语义理解。

图2：训练与推理流程。 （a）训练： 训练期间使用两条路径。上方路径通过线段–圆弧拟合，将干净的 SVG 转换为圆角多边形真值 token。下方路径通过降低分辨率、栅格化及传统矢量化生成退化的轮廓栅格图。VLM 利用交叉熵损失，学习从退化输入中预测干净的几何结构。（b）测试时扩展： 推理时，将输入栅格图随机解码为多个 SVG 候选，对颜色与层叠次序进行后处理，再按 DINO 相似度排序以选出最佳输出（勾号表示最高分）。

图2：训练与推理流程。 （a）训练： 训练期间使用两条路径。上方路径通过线段–圆弧拟合，将干净的 SVG 转换为圆角多边形真值 token。下方路径通过降低分辨率、栅格化及传统矢量化生成退化的轮廓栅格图。VLM 利用交叉熵损失，学习从退化输入中预测干净的几何结构。（b）测试时扩展： 推理时，将输入栅格图随机解码为多个 SVG 候选，对颜色与层叠次序进行后处理，再按 DINO 相似度排序以选出最佳输出（勾号表示最高分）。

Recent data-driven approaches finetune multimodal LLMs to generate Scalable Vector Graphics (SVG) by learning from artist-designed vector data. These methods produce impressive results and are capable of capturing human preferences in control-point placement and layer structuring. However, these approaches face several challenges that limit their robustness and generalizability. First, previous approaches use SVG commands as the underlying representation. While this representation is well suited for visualization, it is neither compact nor canonical from a generative modeling perspective, which complicates the model learning process.

近期的数据驱动方法通过艺术家设计的矢量数据进行学习，对多模态大语言模型微调，以生成 可缩放矢量图形（SVG）。 这些方法取得了令人瞩目的结果，能够捕捉人类在控制点放置和图层组织方面的偏好。 然而，这些方法面临多项挑战，限制了其鲁棒性与泛化能力。 首先，以往方法使用 SVG 命令作为底层表示。 虽然这种表示非常适合可视化，但从生成建模角度看，它既不紧凑，也不规范，这增加了模型学习过程的复杂性。

Second, prior work typically evaluates on clean, noise-free rasters rendered directly from SVG ground truth, but practical image-to-vector tasks demand robustness to diverse input characteristics. As shown in Figure 3, existing methods exhibit unpredictable behavior when the rasterization pipeline is modified—for instance, switching from one rendering backend to another causes significant degradation in output quality. Similarly, outputs from text-to-image models are increasingly used as inputs for vectorization, yet these generated images often contain distorted shapes and visual characteristics that differ substantially from SVG renderings. As a result, previous methods trained on artist-designed vector data struggle to generalize effectively to such real-world scenarios where input characteristics vary. Third, the stochastic nature of autoregressive models can cause them to succeed or fail unpredictably on the same input, resulting in unreliability in practical applications.

其次，以往工作通常在直接由 SVG 真值渲染得到的干净、无噪声栅格图像上进行评估，但实际的图像转矢量任务要求对多样的输入特征保持鲁棒性。 如 图3 所示，当栅格化流程改变时，现有方法会表现出不可预测的行为，例如从一种渲染后端切换到另一种，就会导致输出质量显著下降。 类似地，文生图模型的输出正越来越多地被用作矢量化输入，但这些生成图像通常包含失真的形状，其视觉特征也与 SVG 渲染结果存在显著差异。 因此，在艺术家设计的矢量数据上训练的以往方法，难以有效泛化到输入特征多变的真实世界场景。 第三，自回归模型的随机性可能导致它们面对同一输入时，成功或失败都难以预测，从而使实际应用缺乏可靠性。

图3：对栅格化方法的敏感性。 给定相同的输入 SVG（左侧），当栅格化后端发生变化（CairoSVG 与 Skia）时，StarVector 等现有方法会生成截然不同的输出。这种脆弱性表明，以往工作无法鲁棒应对输入特征多变的实际矢量化场景。

图3：对栅格化方法的敏感性。 给定相同的输入 SVG（左侧），当栅格化后端发生变化（CairoSVG 与 Skia）时，StarVector 等现有方法会生成截然不同的输出。这种脆弱性表明，以往工作无法鲁棒应对输入特征多变的实际矢量化场景。

To address the above-mentioned challenges, we propose a novel LLM-based image-to-vector pipeline that robustly converts images to vector graphics, including those generated by text-to-image models. Our output vector representation is based on a polygonal scaffold with per-vertex radii, providing a compact and canonical form that simplifies model training. Instead of the traditionally used colorful raster inputs, we adopt outline-based raster representations to promote better generalization. Color information is then recovered as a post-processing step by extracting colors from the original input image. During training, we employ a novel synthetic distortion strategy to simulate variations encountered at test time, which further enhances the model’s robustness to generated images. During testing, we introduce a test-time scaling strategy that improves the reliability and consistency of the image-to-vector conversion.

为解决上述挑战，我们提出一种基于 LLM 的新型图像转矢量流程，能够鲁棒地将图像转换为矢量图形，包括由文生图模型生成的图像。 我们的输出矢量表示以多边形骨架为基础，为每个顶点设置半径，提供紧凑且规范的形式，从而简化模型训练。 我们采用基于轮廓的栅格表示，替代传统的彩色栅格输入，以促进更好的泛化。 随后，通过从原始输入图像提取颜色，在后处理步骤中恢复颜色信息。 训练期间，我们采用一种新的合成失真策略来模拟测试时遇到的变化，进一步增强模型对生成图像的鲁棒性。 测试期间，我们引入测试时扩展策略，提高图像转矢量过程的可靠性与一致性。

Contributions. This paper makes several contributions. First, we identify the limited generalization ability of existing VLM-based image vectorization methods and propose a new model that performs robustly in real-world vectorization tasks. Second, we introduce a rounded polygon representation that offers improved geometric accuracy and control-point precision compared to traditional SVG-based formats. Alongside this, we present an outline-based raster representation that remains robust under varying visual characteristics, such as different anti-aliasing methods. Third, we incorporate a novel degradation model during training and apply test-time scaling during inference, which significantly enhance the model’s reliability on imperfect or degraded inputs. Finally, we demonstrate that VectorArk achieves substantially higher fidelity and precision than state-of-the-art methods on both synthetic and real-world datasets.

贡献。 本文具有以下几项贡献。 首先，我们指出现有基于 VLM 的图像矢量化方法泛化能力有限，并提出一种在真实世界矢量化任务中表现鲁棒的新模型。 其次，我们引入圆角多边形表示，相比传统的基于 SVG 的格式，提高了几何准确性与控制点精度。 与此同时，我们提出基于轮廓的栅格表示，使其在不同抗锯齿方式等多变视觉特征下仍保持鲁棒性。 第三，我们在训练期间引入一种新的退化模型，并在推理期间采用测试时扩展，显著增强了模型面对不完美或退化输入时的可靠性。 最后，我们证明 VectorArk 在合成数据集与真实世界数据集上，均能获得显著优于最先进方法的保真度与精度。

2. Related Work

Classical vectorization and differentiable rendering. Classical vectorization methods identify region boundaries by detecting color discontinuities and edge transitions. While computationally efficient, they can be sensitive to anti-aliasing, compression artifacts, and noise. This brittleness is particularly problematic for real-world images such as outputs from text-to-image models or low-resolution graphics where artifacts and distortions cause fragmented, topologically inconsistent boundaries. Moreover, lacking semantic understanding, these methods treat all color transitions equally, resulting in over-segmentation and redundant paths.

传统矢量化与可微渲染。 传统矢量化方法通过检测颜色不连续处和边缘过渡来识别区域边界。 尽管计算效率较高，但它们容易受到抗锯齿、压缩伪影和噪声的影响。 对于文生图模型的输出或低分辨率图形等真实世界图像，这种脆弱性尤其突出，因为其中的伪影和形变会造成破碎且拓扑不一致的边界。 此外，由于缺乏语义理解，这些方法同等对待所有颜色过渡，导致过度分割和冗余路径。

Differentiable rasterization frameworks like DiffVG enable gradient-based optimization of vector parameters, but can produce unstable control points when optimizing for pixel-level reconstruction without geometric priors. Curve regularization techniques still rely on direct edge tracing and lack semantic structure. In contrast, learning-based approaches such as ours encode strong geometric priors from artist-designed data, enabling smooth, semantically coherent primitives that align with human expectations for vector graphic structure.

DiffVG 等可微栅格化框架支持基于梯度优化矢量参数，但如果在缺少几何先验的情况下优化像素级重建，可能产生不稳定的控制点。 曲线正则化技术仍然依赖直接的边缘描摹，缺乏语义结构。 相比之下，我们这类基于学习的方法从艺术家设计的数据中编码了强大的几何先验，从而生成平滑且语义连贯的图元，符合人们对矢量图形结构的预期。

Neural SVG generation with multimodal LMs. Recent approaches leverage large multimodal models for SVG generation. StarVector introduced semantic primitives, and OmniSVG discretized coordinates and commands into tokens; and LLM4SVG integrated SVG tokens into large language models with modular decoders. However, these systems operate within the traditional path-command vocabulary, producing long sequences that amplify small coordinate errors. We demonstrate that a simpler path representation mitigates such error accumulation.

利用多模态语言模型进行神经 SVG 生成。 近期方法使用大型多模态模型生成 SVG。 StarVector 引入了语义图元，OmniSVG 将坐标和命令离散化为 token；LLM4SVG 则借助模块化解码器将 SVG token 整合进大语言模型。 然而，这些系统仍使用传统的路径命令词表，生成的长序列会放大微小的坐标误差。 我们证明，更简单的路径表示能够减轻这种误差累积。

Our design is motivated by the notion of reduced degrees of freedom. Fixing endpoints, a cubic Bézier curve requires four parameters, whereas our representation needs only three. Moreover, our formulation enforces piecewise-constant curvature by construction, yielding smoother and more aesthetically pleasing curves. While recent work on reinforcement learning from rendering feedback has explored improving SVG generation reliability, these methods do not fully eliminate the unpredictable behavior, and effective strategies for ensuring consistent, high-quality vectorization remain an important challenge.

我们的设计源于减少自由度这一思路。 固定端点后，一条三次 Bézier 曲线需要四个参数，而我们的表示只需要三个。 此外，我们的形式化表示在构造上保证了分段恒定的曲率，从而得到更平滑、更美观的曲线。 尽管近期基于渲染反馈的强化学习工作探索了如何提升 SVG 生成的可靠性，但这些方法并未完全消除不可预测的行为；如何确保矢量化结果稳定且质量较高，仍是一项重要挑战。

3. Method

Overview. The input to our model is a raster image, which may be generated by text-to-image models. Our approach aims to automatically convert the input raster into vector graphics while correcting distortions and imperfections present in the original image. The model is finetuned from a pretrained multimodal large language model (Section 3.2). It takes a raster image as input and autoregressively generates a tokenized vector representation.

概述。 模型的输入是一幅栅格图像，也可以由文生图模型生成。 我们的方法旨在自动将输入栅格图转换为矢量图形，同时纠正原始图像中的形变和缺陷。 模型基于预训练的多模态大语言模型进行微调（第 3.2 节）。 它接收栅格图像作为输入，并以自回归方式生成 token 化的矢量表示。

The output vector representation (Section 3.1) is constructed as a scaffold of polygons, where each vertex is assigned a radius that defines an inscribed arc at that polygon vertex. Our input raster employs an outline-based representation (Section 3.1) that is designed to remain robust to discrepancies in visual characteristics between training and testing data. To further enhance robustness in practical image-to-vector tasks, we introduce a degradation model (Section 3.3) during training to simulate visual degradations that may occur during testing, and apply test-time scaling (Section 3.4) to improve the reliability of the pipeline.

输出矢量表示（第 3.1 节）以多边形为骨架构建，为每个顶点指定一个半径，用来定义该顶点处的内切圆弧。 输入栅格图采用基于轮廓的表示（第 3.1 节），旨在应对训练数据与测试数据之间视觉特征的差异。 为进一步增强实际图像矢量化任务中的鲁棒性，我们在训练中引入退化模型（第 3.3 节），模拟测试时可能出现的视觉退化，并采用测试时扩展（第 3.4 节）提升整个流程的可靠性。

3.1. Canonical Representations

(a) Source SVG

(b) Line-arc decomposition

(c) Polygon vertices

图4：圆角多边形表示示例。 从输入 SVG 转换为圆角多边形格式的过程。(a) 具有平滑曲线的原始 SVG。(b) 直线—圆弧分解，其中红色线段表示直线，绿色／蓝色线段表示拟合原始曲线得到的圆弧。(c) 应用圆角半径之前对应的多边形顶点（交点）；最终表示包含这些顶点及编码后的半径。

图4：圆角多边形表示示例。 从输入 SVG 转换为圆角多边形格式的过程。(a) 具有平滑曲线的原始 SVG。(b) 直线—圆弧分解，其中红色线段表示直线，绿色／蓝色线段表示拟合原始曲线得到的圆弧。(c) 应用圆角半径之前对应的多边形顶点（交点）；最终表示包含这些顶点及编码后的半径。

Rounded Polygon Representation. SVGs comprise a wide variety of primitive types—including Bézier curves, circles, ellipses, rectangles, polylines, and more—to represent vector graphics. The heterogeneity poses challenges for training generative models. To simplify the learning process, we adopt a canonical and compact representation that expresses SVGs using rounded polygons. More specifically, we convert each SVG path into a rounded polygon representation independently.

圆角多边形表示。 SVG 使用多种图元表示矢量图形，包括 Bézier 曲线、圆、椭圆、矩形、折线等。 这种异构性给生成模型的训练带来了挑战。 为简化学习过程，我们采用一种规范且紧凑的表示，以圆角多边形表达 SVG。 具体而言，我们将每条 SVG 路径独立转换为圆角多边形表示。

First, we sample equidistant points along the path. Then, we fit line and arc primitives to the sampled points while preserving $G^1$ continuity, following the Cornocopia fitting algorithm. After this step, the SVGs are represented using only two primitives: lines and arcs (see Figure 4b). As shown in our experiments, this conversion yields a more compact representation than the raw SVG representation and is nearly lossless.

首先，我们沿路径采样 等距 点。 然后，遵循 Cornocopia 拟合 算法，对采样点拟合直线和圆弧图元，同时保持 $G^1$ 连续性。 经过这一步，SVG 仅由直线和圆弧两种图元表示（见 图 4b）。 实验表明，与原始 SVG 表示相比，这一转换得到的表示更加紧凑，而且几乎无损。

图5：圆角多边形顶点计算。 给定拟合后的直线与圆弧图元，延长圆弧端点处的切线至其交点，据此计算多边形顶点。

图5：圆角多边形顶点计算。 给定拟合后的直线与圆弧图元，延长圆弧端点处的切线至其交点，据此计算多边形顶点。

In the following step, a rounded polygon representation is derived from the lines and arcs. In this representation, a path is defined as a sequence of triples $\{(x_i,y_i,d_i)\}$, where $(x_i,y_i)$ are the coordinates of the polygon vertices and $d_i$ specifies the roundness at each vertex. For each line primitive, the two endpoints naturally serve as polygon vertices. For each arc primitive, such as the arc $\stackrel{⌢}{DE}$ in Figure 5, the tangent vectors at its endpoints, $\overrightarrow{DB}$ and $\overrightarrow{EB}$, intersect at point $B$.

下一步，从直线和圆弧推导出圆角多边形表示。 在这种表示中，一条路径定义为三元组序列 $\{(x_i,y_i,d_i)\}$，其中 $(x_i,y_i)$ 是多边形顶点的坐标，$d_i$ 指定各顶点的圆角程度。 对于每条直线图元，其两个端点自然成为多边形顶点。 对于每条圆弧图元，例如 图 5 中的圆弧 $\stackrel{⌢}{DE}$，其端点处的切向量 $\overrightarrow{DB}$ 和 $\overrightarrow{EB}$ 相交于点 $B$。

The points $D$, $B$, and $E$ (in order) are taken as the polygon vertices. To simplify the representation, any polygon vertex whose two incident edges are collinear is removed, thereby eliminating redundancy without altering the polygon’s shape. In addition to the coordinates $(x_i,y_i)$, a roundness value $d_i$ is computed for each polygon vertex. For vertices such as $B$, formed by the intersection of arc tangents, the roundness value $d_i$ is defined as the distance $BD$ (see Figure 5).

依次将点 $D$、$B$ 和 $E$ 作为多边形顶点。 为简化表示，移除所有相邻两条边共线的多边形顶点，从而在不改变多边形形状的情况下消除冗余。 除坐标 $(x_i,y_i)$ 外，还为每个多边形顶点计算一个圆角值 $d_i$。 对于 $B$ 这类由圆弧切线交点形成的顶点，圆角值 $d_i$ 定义为距离 $BD$（见 图 5）。

This distance-based parameterization enables more stable quantization than using the arc radius directly, as the radius can become excessively large for nearly flat arcs. The radius of the arc can be computed as $r_i=d_i\tan ({\alpha }_i/2)$, where ${\alpha }_i$ is the interior angle $\mathrm{\angle }DBE$ at the vertex $B$. In this way, the roundness value $d_i$ directly determines the shape of the arc. For other vertices that are not formed by the intersection of arc tangents, the roundness value $d_i$ is set to $-1$, which serves as a special flag indicating that the vertex is an endpoint of a line or an arc.

相比直接使用圆弧半径，这种基于距离的参数化方式能实现更稳定的量化，因为接近直线的圆弧半径可能极大。 圆弧半径可计算为 $r_i=d_i\tan ({\alpha }_i/2)$，其中 ${\alpha }_i$ 是顶点 $B$ 处的内角 $\mathrm{\angle }DBE$。 这样，圆角值 $d_i$ 就直接决定了圆弧的形状。 对于其他并非由圆弧切线交点形成的顶点，将圆角值 $d_i$ 设为 $-1$，作为特殊标记，表示该顶点是直线或圆弧的端点。

We handle important edge cases of arcs as follows. Arcs with negligible curvature are converted to line segments. Large-angle arcs can cause the tangents at their endpoints to intersect far from the arc or not at all; semicircular arcs are a notable example, as their tangents are parallel and never meet. To address this, arcs are uniformly subdivided such that each resulting sub-arc spans an angle of less than ${120}^{\circ }$.

我们按如下方式处理圆弧的重要边界情况。 曲率可以忽略的圆弧被转换为线段。 大角度圆弧可能使端点处的切线在远离圆弧的位置相交，或者根本不相交；半圆弧就是一个典型例子，因为其切线相互平行，永不相交。 为解决这一问题，对圆弧进行均匀细分，使每条子圆弧的张角都小于 ${120}^{\circ }$。

The proposed vector representation employs a rounded polygon format that is both canonical and compact, which simplifies model training, as demonstrated in our experiments.

所提出的矢量表示采用规范且紧凑的圆角多边形格式，能够简化模型训练，实验结果也验证了这一点。

Outline Raster Representation. In image-to-vector conversion tasks, the colors of input rasters can easily exhibit a mismatch between training and testing data. For example, outputs from text-to-image models may have visual characteristics distinct from rendered SVGs, which can hinder the model’s generalization in practical use cases. To address this issue, we provide the multimodal LLM with a “decolorized” raster image as input. More specifically, we first apply a classic vectorization tool to convert the input raster image into a vector graphic.

轮廓栅格表示。 在图像矢量化任务中，训练数据与测试数据的输入栅格图颜色很容易出现不匹配。 例如，文生图模型的输出可能具有不同于 SVG 渲染图的视觉特征，这会妨碍模型在实际场景中的泛化。 为解决这一问题，我们向多模态大语言模型提供“去色”的栅格图像作为输入。 具体而言，首先使用传统矢量化工具，将输入栅格图像转换为矢量图形。

We then discard its color information and render only the vector paths as a black-and-white raster outline with fixed stroke-width (see Figure 2). Although traditional vectorization tools may produce suboptimal vector geometries, they can still faithfully reconstruct the original image. Since our method only requires the rendered output of the vectorization, this step effectively “decolorizes” the input image while preserving its geometric structure. This procedure can be interpreted as a normalization step that maps each input image to a canonical view, thereby improving the model’s robustness to out-of-distribution test data.

然后丢弃其颜色信息，仅将矢量路径渲染为描边宽度固定的黑白栅格轮廓图（见 图 2）。 虽然传统矢量化工具生成的矢量几何形状可能不够理想，但仍能忠实重建原始图像。 由于我们的方法只需要矢量化结果的渲染输出，这一步能够有效地对输入图像“去色”，同时保留其几何结构。 这一过程可以视为归一化步骤，将每幅输入图像映射到规范视图，从而提升模型对分布外测试数据的鲁棒性。

Our multimodal LLM is trained to predict colorless outputs from colorless inputs, which greatly simplifies the training process and enhances generalization. During inference, we apply a post-processing step based on a path color extraction method to recover the colors and z-order of output vector paths from the original input image (see the supplementary material for details). As demonstrated in our experiments, the outline raster representation yields better results than the color raster input, and the color extraction method effectively propagates the input colors to the output vector graphics.

我们的多模态大语言模型学习从无色输入预测无色输出，大幅简化了训练过程并增强了泛化能力。 推理时，采用基于路径颜色提取方法的后处理步骤，从原始输入图像中恢复输出矢量路径的颜色和层叠次序（详见补充材料）。 实验表明，轮廓栅格表示比彩色栅格输入取得更好的结果，颜色提取方法也能有效地将输入颜色传递到输出矢量图形中。

3.2. Sequence Modeling with Multimodal LLMs

We encode our rounded polygon representation as a text sequence. Our model is finetuned from a pretrained multimodal LLM (InternVL in our experiments) that takes an image as input and generates the text tokens in an autoregressive manner. More specifically, we normalize all vector graphics to a $128\times 128$ viewBox with a 4 unit padding on each side. The coordinates $(x_i,y_i)$ and roundness parameters $d_i$ of the target rounded polygon representation are quantized to two decimal places.

我们将圆角多边形表示编码为文本序列。 模型基于预训练的多模态大语言模型进行微调（实验中使用 InternVL），接收图像输入，并以自回归方式生成文本 token。 具体而言，将所有矢量图形归一化到 $128\times 128$ 的 viewBox 中，并在四周各留出 4 个单位的边距。 目标圆角多边形表示中的坐标 $(x_i,y_i)$ 和圆角参数 $d_i$ 均量化到小数点后两位。

The sequence $(x_i,y_i,d_i)$ is then serialized into text, where each triplet component $x_i,y_i,d_i$ is separated by [,], consecutive polygon vertices are separated by [space], and different paths are separated by [\n]. We reuse the tokenizer vocabulary from the pretrained multimodal LLM backbone, as this leads to more stable and efficient convergence compared to introducing new tokens. Following InternVL, the input image resolution is set to $448\times 448$. The image is processed by the ViT encoder to produce visual tokens, which are then used as conditioning inputs to the LLM backbone.

随后将序列 $(x_i,y_i,d_i)$ 序列化为文本，其中每个三元组的分量 $x_i,y_i,d_i$ 由 [,] 分隔，相邻多边形顶点由 [space] 分隔，不同路径由 [\n] 分隔。 我们复用预训练多模态大语言模型主干的分词器词表，因为相比引入新 token，这样能实现更稳定、更高效的收敛。 遵循 InternVL，将输入图像分辨率设为 $448\times 448$。 图像经过 ViT 编码器处理后产生视觉 token，随后作为条件输入送入大语言模型主干。

During finetuning, we update all parameters of the multimodal LLM, including those of the ViT encoder. We observe that finetuning only the LLM backbone or applying LoRA leads to inferior performance. We adopt the next-token prediction objective, using cross-entropy loss over all text tokens to supervise training.

微调过程中，我们更新多模态大语言模型的全部参数，包括 ViT 编码器的参数。 我们观察到，仅微调大语言模型主干或使用 LoRA 的性能较差。 我们采用 下一 token 预测 目标，通过所有文本 token 上的交叉熵损失监督训练。

3.3. Degradation Model

Our model is trained on a dataset of high-quality vector graphics. A straightforward way to prepare training data is to render each vector graphic into a high-fidelity raster image and train the model to reconstruct the original vector from this raster input. However, this approach performs poorly when the input rasters contain degraded or distorted shapes and geometry—such as those produced by text-to-image models. In such cases, the model tends to faithfully reproduce the imperfections of the input raster, leading to low-quality vector graphics. Another approach is to introduce random noise to the control points of the outline representation described in Section 3.2 before rendering it into raster inputs. However, we found that this method does not generalize well to practical image-to-vector tasks.

模型使用高质量矢量图形数据集训练。 准备训练数据的一种直接方法，是将每幅矢量图形渲染为高保真栅格图像，再训练模型从该栅格输入重建原始矢量图。 然而，当输入栅格图包含退化或扭曲的形状与几何结构时，例如文生图模型生成的图像，这种方法的表现较差。 在这些情况下，模型往往会忠实复现输入栅格图中的缺陷，导致矢量图形质量低下。 另一种方法是在将 第 3.2 节 所述的轮廓表示渲染为栅格输入之前，向其控制点添加随机噪声。 但我们发现，这种方法无法很好地泛化到实际图像矢量化任务。

We propose a vectorization-based degradation model that effectively simulates the types of imperfections commonly observed in practical scenarios. Specifically, for each training vector graphics, we render it at a randomly sampled reduced resolution (ranging from $224\times 224$ to $336\times 336$ in our experiments) and apply a random Gaussian blur. Since classical vectorizers are sensitive to raster resolution, lower-resolution inputs tend to produce poorer vectorization results. We leverage this property by applying a classical vectorizer to obtain an imperfect outline, which is then rendered back into a raster image and used as the model’s input during training. As demonstrated in our experiments, the proposed degradation model enables our method to generalize effectively across diverse image inputs, including those generated by text-to-image models.

我们提出一种基于矢量化的退化模型，能够有效模拟实际场景中常见的各类缺陷。 具体而言，对每幅训练矢量图形，以随机采样的较低分辨率进行渲染（实验中从 $224\times 224$ 到 $336\times 336$），并施加随机高斯模糊。 由于传统矢量化工具对栅格分辨率敏感，较低分辨率的输入通常会产生较差的矢量化结果。 我们利用这一特性，使用传统矢量化工具获得不完美的轮廓，再将其重新渲染为栅格图像，作为训练时的模型输入。 实验表明，所提出的退化模型使我们的方法能够有效泛化到多种图像输入，包括文生图模型生成的图像。

3.4. Test-Time Scaling

Compared to classical vectorizers, LLM-based image-to-vector methods can succeed or fail unpredictably on the same input when evaluated with different random seeds. To address this issue, we apply test-time scaling during inference. For each input image, we generate $N$ candidate vector graphics through stochastic decoding of our image-to-vector model. We then recover the colors and z-order of the output vector paths from the input raster.

相比传统矢量化工具，基于大语言模型的图像矢量化方法在使用不同随机种子进行评估时，可能对同一输入时而成功、时而失败，难以预测。 为解决这一问题，我们在推理过程中采用测试时扩展。 对于每幅输入图像，通过对图像矢量化模型进行随机解码，生成 $N$ 个候选矢量图形。 然后从输入栅格图中恢复输出矢量路径的颜色和层叠次序。

Each candidate vector graphic is rendered back into a raster image, and both these rendered images and the original input raster are encoded into embeddings using the frozen DINO-ViT-B/16 encoder. The final output is selected as the candidate whose rendered raster has the highest cosine similarity with the input raster in the embedding space. Among various feature extractors we tested, DINO yielded the most reliable results, consistent with observations from StarVector, significantly improving the robustness of our pipeline, which is crucial for practical image-to-vector applications.

将每个候选矢量图形重新渲染为栅格图像，并使用冻结的 DINO-ViT-B/16 编码器，将这些渲染图像与原始输入栅格图一同编码为嵌入。 最终选择其渲染栅格图与输入栅格图在嵌入空间中余弦相似度最高的候选作为输出。 在测试的多种特征提取器中，DINO 的结果最为可靠，这与 StarVector 的观察一致，显著提升了整个流程的鲁棒性，而这对实际图像矢量化应用至关重要。

4. Experiments

4.1. Experimental Setup

Implementation Details. Our model is finetuned from InternVL2-1B, in an end-to-end manner. The ViT encoder processes outline rasters at $448\times 448$ resolution. We employ AdamW optimizer with cosine learning rate decay, initial learning rate ${10}^{-4}$, batch size 256, and 250K iterations. Training uses $\sim 5M$ SVGs from icons, logos, and flat graphics, with random rotation and scaling augmentation. The degradation model is dropped with probability 25% to ensure robustness on clean inputs.

实现细节。 我们以端到端方式对 InternVL2-1B 进行微调，得到本文模型。 ViT 编码器处理分辨率为 $448\times 448$ 的轮廓栅格图。 我们使用 AdamW 优化器和余弦学习率衰减，初始学习率为 ${10}^{-4}$，批大小为 256，训练 250K 次迭代。 训练使用约 500 万幅图标、标志和平面图形 SVG，并进行随机旋转和缩放增强。 以 25% 的概率跳过退化模型，以保证对干净输入的鲁棒性。

During training, we normalize the ground-truth SVGs to a $128\times 128$ viewBox and apply a classic vectorizer to create stroke-only outlines. During inference, given a raster input, we apply the same vectorization pipeline and parameters used during training to extract a stroke-only outline. This acts as a “normalization” step to avoid training-testing mismatch.

训练期间，将真实 SVG 归一化到 $128\times 128$ 的 viewBox 中，并使用传统矢量化工具生成仅含描边的轮廓。 推理时，对给定的栅格输入，采用与训练时 相同 的矢量化流程和参数，提取仅含描边的轮廓。 这相当于一个“归一化”步骤，用于避免训练与测试之间的不匹配。

Benchmark Datasets. We benchmark our method on the two latest SVG generation benchmarks: SArena and SVGenius. To facilitate systematic evaluation across varying levels of geometric complexity, we stratify samples into difficulty tiers following SVGenius. For SArena, we apply a primitive-count-based categorization (Easy: <64, Medium: 64–128, Hard: >128 path primitives), while for SVGenius, we adopt their predefined complexity levels.

基准数据集。 我们在两个最新的 SVG 生成基准 SArena 和 SVGenius 上评估所提出的方法。 为系统评估不同几何复杂度下的表现，我们遵循 SVGenius，将样本划分为不同难度等级。 对于 SArena，按路径图元数量进行划分（简单：<64，中等：64–128，困难：>128 个路径图元）；对于 SVGenius，则采用其预定义的复杂度等级。

Evaluation Metrics. Following OmniSVG, we assess reconstruction quality using four metrics: LPIPS, SSIM, DINO (ViT-B/16 cosine similarity), and MSE. These metrics jointly capture perceptual, structural, and pixel-wise fidelity of generated SVGs relative to ground truth.

评估指标。 遵循 OmniSVG，我们使用四项指标评估重建质量：LPIPS、SSIM、DINO（ViT-B/16 余弦相似度）和 MSE。 这些指标共同衡量生成的 SVG 相对于真实值在感知、结构和像素层面的保真度。

表1：SArena 和 SVGenius 基准上各难度等级的定量比较。我们的方法在所有指标和复杂度类别上均取得最佳表现。粗体表示各指标的最佳分数。

SArena
Diff./Metric	GPT-4o	Gemini	OmniSVG	StarVector	Ours
E / SSIM ↑	0.681	0.622	0.823	0.876	0.937
E / LPIPS ↓	0.205	0.253	0.099	0.069	0.031
E / MSE ↓	0.095	0.121	0.036	0.032	0.011
E / Dino ↑	0.972	0.944	0.98	0.969	0.992
M / SSIM ↑	0.53	0.493	0.6	0.75	0.895
M / LPIPS ↓	0.284	0.323	0.251	0.142	0.058
M / MSE ↓	0.133	0.163	0.125	0.062	0.013
M / Dino ↑	0.958	0.932	0.903	0.949	0.981
H / SSIM ↑	0.47	0.441	0.518	0.626	0.857
H / LPIPS ↓	0.334	0.382	0.324	0.252	0.093
H / MSE ↓	0.151	0.169	0.123	0.101	0.022
H / Dino ↑	0.931	0.897	0.898	0.902	0.975

SVGenius
Diff./Metric	GPT-4o	Gemini	OmniSVG	StarVector	Ours
E / SSIM ↑	0.673	0.611	0.84	0.89	0.944
E / LPIPS ↓	0.19	0.244	0.07	0.046	0.028
E / MSE ↓	0.094	0.121	0.027	0.019	0.008
E / Dino ↑	0.976	0.951	0.985	0.993	0.995
M / SSIM ↑	0.572	0.539	0.674	0.71	0.868
M / LPIPS ↓	0.28	0.328	0.204	0.203	0.08
M / MSE ↓	0.087	0.097	0.058	0.058	0.015
M / Dino ↑	0.942	0.914	0.94	0.921	0.977
H / SSIM ↑	0.566	0.536	0.638	0.672	0.83
H / LPIPS ↓	0.295	0.326	0.248	0.258	0.12
H / MSE ↓	0.079	0.092	0.065	0.059	0.023
H / Dino ↑	0.928	0.918	0.918	0.893	0.958

表1：SArena 和 SVGenius 基准上各难度等级的定量比较。我们的方法在所有指标和复杂度类别上均取得最佳表现。粗体表示各指标的最佳分数。

4.2. Comparison with Prior Art

For comparative evaluation, we include recent state-of-the-art multimodal baselines: OmniSVG, StarVector, as well as proprietary multimodal models GPT-4o and Gemini. For OmniSVG and StarVector, we use their models with highest performance. For GPT-4o and Gemini, we adopt the prompts from SVGenius. All baselines employ identical test-time scaling configurations to ensure fair comparison.

为进行比较评估，我们纳入近期先进的多模态基线 OmniSVG、StarVector，以及专有多模态模型 GPT-4o 和 Gemini。 对于 OmniSVG 和 StarVector，我们使用其性能最佳的模型。 对于 GPT-4o 和 Gemini，我们采用 SVGenius 中的提示词。 所有基线均使用相同的测试时扩展配置，以保证公平比较。

Quantitative Evaluation. Quantitative results are summarized in Table 1. Although our model has fewer parameters compared to the baselines, we observe a consistent and significant improvement across all four metrics, LPIPS, SSIM, DINO, and MSE across both SArena and SVGenius benchmarks, with particularly pronounced gains on higher complexity tiers. This advantage persists even in the single-sample setting without test-time scaling ($N=1$) on SVGenius, where our method still outperforms the strongest baseline.

定量评估。 定量结果汇总于 表 1。 尽管模型参数量少于各基线，但在 SArena 和 SVGenius 两个基准上，我们均观察到 LPIPS、SSIM、DINO 和 MSE 四项指标持续且显著的提升，其中在高复杂度等级上的增益尤其明显。 即便在 SVGenius 上采用不进行测试时扩展的单样本设置（$N=1$），这一优势依然存在，我们的方法仍优于最强基线。

Input Raster

StarVector 8B

OmniSVG

GPT-4o

Ours

图6：对文生图模型生成的输入进行定性比较。

图6：对文生图模型生成的输入进行定性比较。

Qualitative Evaluation. Figure 6 presents qualitative comparisons on challenging inputs generated by text-to-image models (GPT-4o and Gemini). These generated images often contain artifacts and distorted shapes that substantially degrade vectorization quality. While all baselines struggle with such imperfect inputs and produce imprecise geometry, our approach demonstrates significantly superior robustness.

定性评估。 图 6 展示了针对文生图模型（GPT-4o 和 Gemini）生成的具有挑战性的输入所进行的定性比较。 这些生成图像通常包含伪影和扭曲形状，会显著降低矢量化质量。 各基线都难以处理此类不完美输入，生成的几何形状不够精确，而我们的方法表现出显著更强的鲁棒性。

4.3. Ablation Study

Ablation on Rounded Polygon Representation We analyze the impact of our proposed rounded polygon representation by keeping the dataset, model architecture, and training setup identical, while substituting our representation with OmniSVG's and StarVector's representation. For the OmniSVG representation, all SVG commands are restricted to MoveTo (M), LineTo (L), Arc (A), CubicBézier (C), and ClosePath (Z) using only absolute coordinates. We enforce a fixed viewBox with standardized padding, and quantize all coordinates to two decimal places of precision following our setup. All SVG constructs—regardless of their original type—are expanded into <path> elements, and their corresponding “d” attributes are extracted and concatenated into a single string representation, with individual paths separated by newline characters. For the StarVector representation, we adopt their native SVG format which uses the input raw SVG structure.

圆角多边形表示的消融实验 我们保持数据集、模型架构和训练设置不变，将所提出的表示分别替换为 OmniSVG 和 StarVector 的表示，以分析圆角多边形表示的影响。 对于 OmniSVG 表示，将所有 SVG 命令限制为 MoveTo（M）、LineTo（L）、Arc（A）、CubicBézier（C）和 ClosePath（Z），且仅使用绝对坐标。 我们固定 viewBox 并采用标准化边距，遵循本文设置，将全部坐标量化到小数点后两位。 无论原始类型为何，都将所有 SVG 构造展开为 <path> 元素，提取对应的“d”属性并连接成单个字符串表示，各条路径之间用换行符分隔。 对于 StarVector 表示，我们采用其原生 SVG 格式，直接使用输入的原始 SVG 结构。

表2：在 SVGenius 数据集的不同难度等级上，比较 OmniSVG 表示、StarVector 表示与我们提出的直线—圆弧混合表示的消融实验。粗体表示各指标的最佳分数。

Difficulty	Metric	OmniSVG	StarVector	Ours
Easy	SSIM ↑	0.987	0.95	0.944
Easy	LPIPS ↓	0.01	0.029	0.028
Easy	MSE ↓	0.004	0.01	0.008
Easy	DinoScore ↑	0.994	0.992	0.995
Medium	SSIM ↑	0.858	0.697	0.868
Medium	LPIPS ↓	0.111	0.242	0.08
Medium	MSE ↓	0.025	0.058	0.015
Medium	DinoScore ↑	0.959	0.908	0.977
Hard	SSIM ↑	0.743	0.628	0.83
Hard	LPIPS ↓	0.204	0.31	0.12
Hard	MSE ↓	0.04	0.066	0.023
Hard	DinoScore ↑	0.923	0.866	0.958

表2：在 SVGenius 数据集的不同难度等级上，比较 OmniSVG 表示、StarVector 表示与我们提出的直线—圆弧混合表示的消融实验。粗体表示各指标的最佳分数。

The quantitative evaluation is presented in Table 2.

定量评估见 表 2。

Our rounded-polygon representation substantially outperforms the baselines on the medium and hard splits, highlighting the benefit of the proposed canonical representation. On the Easy split, the raw SVG-command representation attains slightly stronger pixel-level metrics while our representation leads in semantic similarity (DINO); we analyze this behavior in detail in the supplementary material.

我们的圆角多边形表示在中等和困难子集上显著优于基线，凸显了所提出的规范表示的优势。 在简单子集上，原始 SVG 命令表示的像素级指标略优，而我们的表示在语义相似度（DINO）方面领先；补充材料对此进行了详细分析。

表3：各难度等级的 token 效率比较。 我们的表示在保持较高重建质量（所有 DINO 分数 > 0.99）的同时，显著减少了 token 数量（27.9–46.6%）。

Dataset/Difficulty	Ours	OmniSVG	Savings (%)
SArena Icon Generation
Easy	642	890	27.9
Medium	1568	2332	32.7
Hard	3130	4637	32.5
SVGenius
Easy	613	960	36.1
Medium	2694	5046	46.6
Hard	4345	7838	44.6

表3：各难度等级的 token 效率比较。 我们的表示在保持较高重建质量（所有 DINO 分数 > 0.99）的同时，显著减少了 token 数量（27.9–46.6%）。

Efficiency of the Rounded Polygon Representation. Our representation is also more token-efficient. As shown in Table 3, we achieve consistent token reductions ranging from 27.9% to 46.6% across all difficulty levels on both benchmarks while maintaining nearly lossless reconstruction quality (DINO score > 0.99). This directly reduces memory footprint and inference cost. In practice, single-SVG generation takes only 3–4 s on an A100.

圆角多边形表示的效率。 我们的表示也具有更高的 token 效率。 如 表 3 所示，在两个基准的所有难度等级上，token 数量均减少了 27.9% 至 46.6%，同时保持几乎无损的重建质量（DINO 分数 > 0.99）。 这直接降低了显存占用和推理成本。 实际使用中，在 A100 上生成单幅 SVG 仅需 3–4 秒。

Ablation on Degradation Model To assess the effectiveness of our degradation model, we conduct an ablation study comparing our full model against an identical configuration trained without the degradation pipeline. All test inputs are generated by text-to-image models (GPT-4o and Gemini), representing practical deployment scenarios where raster graphics often contain artifacts and imperfections. As illustrated in Figure 7, our degradation pipeline enables the model to produce significantly cleaner and more regularized geometry compared to both traditional vectorization (Adobe Illustrator's Image Trace) and the baseline variant trained without training-time degradation simulation. The model trained without degradation model tends to reproduce input artifacts, limiting geometric quality and editability, whereas our approach demonstrates robust artifact suppression and improved structural coherence.

退化模型的消融实验 为评估退化模型的有效性，我们将完整模型与配置相同、但训练时不使用退化流程的模型进行消融比较。 所有测试输入均由文生图模型（GPT-4o 和 Gemini）生成，以代表栅格图形经常包含伪影和缺陷的实际部署场景。 如 图 7 所示，相比传统矢量化方法（Adobe Illustrator 的图像描摹功能）和训练时不模拟退化的基线变体，我们的退化流程使模型能够生成显著更干净、更规整的几何形状。 不使用退化模型训练的模型往往会复现输入伪影，限制了几何质量和可编辑性，而我们的方法能稳健抑制伪影并提升结构连贯性。

图7：使用与不使用退化模型的质量比较。 四个测试案例（以 2×2 网格排列）比较不同方法的矢量化质量。对于每幅输入图像，展示以下方法生成的轮廓：Adobe Illustrator（图像描摹）、使用退化模型的本文方法 以及 不使用退化模型的本文方法。Adobe Illustrator 会引入伪影和不规则几何形状（方框标出）。通过噪声感知训练，我们的方法显著减轻了这些伪影，生成更干净、规整且适合编辑的几何形状。没有此类训练时，模型往往会复现输入噪声和伪影，无法达到相当的几何质量与可编辑性。尤其是在水枪示例中，我们的方法重建了平滑曲线，并将复杂外壳分解为更小、更易编辑的组件。

图7：使用与不使用退化模型的质量比较。 四个测试案例（以 2×2 网格排列）比较不同方法的矢量化质量。对于每幅输入图像，展示以下方法生成的轮廓：Adobe Illustrator（图像描摹）、使用退化模型的本文方法 以及 不使用退化模型的本文方法。Adobe Illustrator 会引入伪影和不规则几何形状（方框标出）。通过噪声感知训练，我们的方法显著减轻了这些伪影，生成更干净、规整且适合编辑的几何形状。没有此类训练时，模型往往会复现输入噪声和伪影，无法达到相当的几何质量与可编辑性。尤其是在水枪示例中，我们的方法重建了平滑曲线，并将复杂外壳分解为更小、更易编辑的组件。

图8：SVGenius 上的测试时扩展分析。 DinoScore 随并行生成数量的变化，取三个随机种子的平均结果。增加并行样本能够通过多样性采样提升重建质量。

图8：SVGenius 上的测试时扩展分析。 DinoScore 随并行生成数量的变化，取三个随机种子的平均结果。增加并行样本能够通过多样性采样提升重建质量。

Ablation on Test-Time Scaling We investigate the effect of test-time scaling by varying the number of parallel generations and selecting the best output based on DinoScore. For each configuration, we generate $n$ samples in parallel and select the highest-quality output. As shown in Figure 8, the quality improves with the number of parallel generations, demonstrating the benefit of test-time scaling. While additional samples continue to improve quality, the gains become marginal beyond a certain threshold, indicating diminishing returns relative to computational cost. This trade-off allows users to balance quality requirements against inference budget in deployment scenarios.

测试时扩展的消融实验 我们改变并行生成数量，并根据 DinoScore 选择最佳输出，以研究测试时扩展的影响。 对于每种配置，并行生成 $n$ 个样本，再选择质量最高的输出。 如 图 8 所示，质量随并行生成数量增加而提升，验证了测试时扩展的收益。 尽管增加样本仍会改善质量，但超过一定阈值后增益趋于微小，表明相对于计算成本，其收益逐渐递减。 这一权衡使用户能够在部署场景中平衡质量要求与推理预算。

Ablation on Outline Raster Representation. We compare model performance when conditioned on colored versus outline-only inputs. Using identical architecture and training configuration, we train another variant on RGB inputs. Quantitative evaluation is presented in Table 4. Outline conditioning consistently outperforms color-based training across all metrics, benchmarks, and difficulty levels, with particularly pronounced gains on harder samples. Pre-vectorizing the input to extract outlines acts as an effective normalization step, eliminating appearance-related training-testing mismatch and allowing the model to focus on geometric structure. This leads to improved shape coherence and topological consistency in the generated SVGs.

轮廓栅格表示的消融实验。 我们比较了以 彩色 输入和 仅含轮廓 的输入作为条件时的模型性能。 采用相同的架构和训练配置，另行训练一个以 RGB 图像为输入的变体。 定量评估见 表 4。 在所有指标、基准和难度等级上，轮廓条件始终优于基于颜色的训练，在更困难的样本上提升尤为明显。 预先对输入进行矢量化以提取轮廓，相当于有效的归一化步骤，消除了外观引起的训练与测试不匹配，使模型能够专注于几何结构。 这提升了生成 SVG 的形状连贯性和拓扑一致性。

表4：在相同模型架构和训练设置下，比较彩色输入与仅含轮廓的输入条件的消融实验。结果涵盖 SArena 和 SVGenius 两个基准的三个难度等级。粗体表示更好的性能。

Dataset/Difficulty	SSIM ↑	LPIPS ↓	MSE ↓	Dino ↑
SArena Icon Generation
Easy (Color)	0.878	0.055	0.031	0.978
Easy (Outline)	0.937	0.031	0.011	0.992
Medium (Color)	0.793	0.091	0.023	0.953
Medium (Outline)	0.895	0.058	0.013	0.981
Hard (Color)	0.715	0.162	0.033	0.928
Hard (Outline)	0.857	0.093	0.022	0.975
SVGenius
Easy (Color)	0.906	0.061	0.024	0.983
Easy (Outline)	0.944	0.028	0.008	0.995
Medium (Color)	0.761	0.118	0.023	0.947
Medium (Outline)	0.868	0.08	0.015	0.977
Hard (Color)	0.697	0.165	0.032	0.887
Hard (Outline)	0.83	0.12	0.023	0.958

表4：在相同模型架构和训练设置下，比较彩色输入与仅含轮廓的输入条件的消融实验。结果涵盖 SArena 和 SVGenius 两个基准的三个难度等级。粗体表示更好的性能。

5. Conclusion

We introduced a geometry-first VLM pipeline for raster-to-SVG vectorization, combining a novel rounded polygon representation, a new degradation model, and a test-time scaling strategy. Across benchmarks and difficult real-world inputs, the method yields compact, faithful, and robust SVGs. Ablations clarify why each component matters, offering practical guidance for vector graphics generative models. Our representation targets medium-complexity graphics; highly detailed illustrations and text/gradient effects are simplified, and dense local path structure remains challenging (see supplementary). Future work includes scaling to richer primitives under controlled constraints and integrating a learned appearance module while preserving geometric fidelity.

我们提出了一个以几何为先的 VLM 栅格图到 SVG 矢量化流程，结合新颖的圆角多边形表示、新的退化模型和测试时扩展策略。 在各基准和具有挑战性的真实世界输入上，该方法均能生成紧凑、忠实且稳健的 SVG。 消融实验阐明了各组件发挥作用的原因，为矢量图形生成模型提供了实用指导。 我们的表示面向 中等复杂度 图形；高度精细的插图、文本及渐变效果会被简化，密集的局部路径结构仍具有挑战性（见补充材料）。 未来工作包括在可控约束下扩展到更丰富的图元，以及在保持几何保真度的同时整合可学习的外观模块。