Lyirs的小屋

主题

InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models ​

SVGMLLMUnified ModelingICLR 2026上海交通大学上海人工智能实验室南京大学东华大学香港中文大学

Wang H, Yin J, Wei Q, et al. InternSVG: towards unified SVG tasks with multimodal large language models[C/OL]//The Fourteenth International Conference on Learning Representations. 2026[2026-10-25].

https://arxiv.org/abs/2510.11341v4

https://hmwang2002.github.io/release/internsvg/

https://github.com/hmwang2002/InternSVG

InternSVG：利用多模态大语言模型统一 SVG 任务 ​

InternSVG 系列概览。SAgoge 提供跨多个领域的大规模、多样化 SVG 样本。SArena 支持对现有 MLLM 的 SVG 任务能力进行全面评估。InternSVG 支持 SVG 理解、编辑和生成的统一建模。

图1：InternSVG 系列概览。SAgoge 提供跨多个领域的大规模、多样化 SVG 样本。SArena 支持对现有 MLLM 的 SVG 任务能力进行全面评估。InternSVG 支持 SVG 理解、编辑和生成的统一建模。

图1：InternSVG 系列概览。SAgoge 提供跨多个领域的大规模、多样化 SVG 样本。SArena 支持对现有 MLLM 的 SVG 任务能力进行全面评估。InternSVG 支持 SVG 理解、编辑和生成的统一建模。

Abstract ​

General SVG modeling remains challenging due to fragmented datasets, limited transferability of methods across tasks, and the difficulty of handling structural complexity. In response, we leverage the strong transfer and generalization capabilities of multimodal large language models (MLLMs) to achieve unified modeling for SVG understanding, editing, and generation. We present the InternSVG family, an integrated data–benchmark–model suite. At its core is SAgoge, the largest and most comprehensive multimodal dataset for SVG tasks, encompassing both static graphics and dynamic animations. It covers icons, long-sequence illustrations, scientific diagrams, and dynamic animations, supporting tasks of varied difficulty levels and providing deeper hierarchies with richer attributes compared to previous datasets. Based on this resource, we introduce SArena, a companion benchmark with comprehensive task definitions and standardized evaluation that aligns with the domains and difficulty spectrum covered by SAgoge. Building on these foundations, we propose InternSVG, a unified MLLM for SVG understanding, editing, and generation with SVG-specific special tokens, subword-based embedding initialization, and a two-stage training strategy that progresses from short static SVGs to long-sequence illustrations and complex animations. This unified formulation induces positive transfer and improves overall performance. Experiments on SArena and prior benchmark confirm that InternSVG achieves substantial gains and consistently outperforms leading open and proprietary counterparts.

由于数据集分散、方法在任务之间的迁移能力有限，以及结构复杂性难以处理，通用 SVG 建模仍面临挑战。 为此，我们利用多模态大语言模型（MLLM）的强大迁移与泛化能力，对 SVG 理解、编辑和生成进行统一建模。 我们提出 InternSVG 系列，一个集数据、基准与模型于一体的体系。 其核心是 SAgoge，这是规模最大、覆盖最全面的多模态 SVG 任务数据集，同时包含静态图形与动态动画。 它涵盖图标、长序列插图、科学图示和动态动画，支持不同难度的任务，与以往数据集相比，具有更深的层级和更丰富的属性。 基于这一资源，我们推出配套基准 SArena，提供全面的任务定义和标准化评估，与 SAgoge 覆盖的领域和难度范围相对应。 在此基础上，我们提出 InternSVG，一个用于 SVG 理解、编辑和生成的统一 MLLM，引入 SVG 专用特殊 token、基于子词的嵌入初始化，以及从短序列静态 SVG 逐步过渡到长序列插图和复杂动画的两阶段训练策略。 这种统一建模方式促进了正迁移，并提升了整体性能。 在 SArena 和已有基准上的实验证实，InternSVG 取得了显著提升，持续优于领先的开放模型和专有模型。

1. Introduction ​

Scalable Vector Graphics (SVG) is an XML-based standard for 2D graphics that offers compact storage, fine-grained editability, and resolution-independent rendering across displays. It can function either as a stand-alone format or as an embedded component within other documents, while also supporting interactivity and animation. Additionally, it interoperates smoothly with CSS, the DOM, and JavaScript in web contexts. These properties have led to broad adoption in web design, scientific visualization, and computer-aided design. Nevertheless, enabling machines to understand, generate, and edit SVGs remains challenging due to the methodological complexity involved, the scarcity of large-scale and high-quality training corpora, and the requirement for models that generalize across tasks rather than specializing in isolation.

可缩放矢量图形（SVG）是一种基于 XML 的二维图形标准，具有存储紧凑、可精细编辑以及跨显示设备渲染时与分辨率无关等特点。 它既可作为独立格式使用，也可作为组件嵌入其他文档，同时支持交互和动画。 此外，它能在 Web 环境中与 CSS、DOM 和 JavaScript 顺畅互操作。 这些特性使其广泛应用于网页设计、科学可视化和计算机辅助设计。 然而，由于相关方法复杂、大规模高质量训练语料稀缺，且模型需要具备跨任务泛化能力而非仅专精于单一任务，让机器理解、生成和编辑 SVG 仍然具有挑战性。

Existing resources and methods face three major limitations. First, current datasets provide limited support for unified modeling of SVG tasks. Most datasets and benchmarks target a single task, such as semantic understanding in SGP-Bench, instruction-based editing in SVGEditBench, or Text-to-SVG and Image-to-SVG generation in ColorSVG-100K, SVG-Stack, and MMSVG. This narrow scope leads to fragmented supervision and evaluation, which in turn limits transfer across understanding, editing, and generation.

现有资源与方法面临三项主要局限。 首先，当前数据集对 SVG 任务统一建模的支持有限。 多数数据集和基准面向单一任务，例如 SGP-Bench 中的语义理解、SVGEditBench 中的指令式编辑，或 ColorSVG-100K、SVG-Stack 和 MMSVG 中的文本到 SVG 与图像到 SVG 生成。 这种狭窄的范围造成了监督与评估的割裂，进而限制了理解、编辑和生成任务之间的迁移。

Second, existing datasets are limited in both scale and diversity. UniSVG focuses on unified SVG understanding and generation, but it contains only about 525k samples. SVGenius provides a comprehensive benchmark spanning understanding, editing, and generation, while it contains only about 2,400 queries, making it suitable for evaluation rather than large-scale training. More broadly, current datasets concentrate on common static images such as icons and illustrations, while paying insufficient attention to SVGs with specific applications in professional domains.

其次，现有数据集在规模和多样性方面均存在不足。 UniSVG 关注统一的 SVG 理解与生成，但仅包含约 52.5 万个样本。 SVGenius 提供了涵盖理解、编辑和生成的综合基准，但仅包含约 2,400 条查询，因此适合评估，却不足以支撑大规模训练。 更广泛地看，当前数据集主要关注图标和插图等常见静态图像，对专业领域中特定应用的 SVG 关注不足。

Third, existing methods commonly lack transferability and generalization. Optimization-based and differentiable rasterization pipelines scale poorly and lack semantic reasoning. Although some approaches employ diffusion models to improve visual fidelity, the resulting SVGs often suffer from limited editability and weak detail-aware control. Recently, LLM-based methods have achieved significant progress in Text-to-SVG and Image-to-SVG generation, but they struggle to generalize to long sequences and complex SVG content, making it difficult to ensure generation quality. Moreover, these methods largely overlook tasks related to SVG understanding and editing.

第三，现有方法普遍缺乏迁移能力与泛化能力。 基于优化和可微栅格化的流程扩展能力较差，且缺乏语义推理。 尽管部分方法采用扩散模型提升视觉保真度，但所得 SVG 往往可编辑性有限，对细节的控制也较弱。 近期，基于 LLM 的方法在文本到 SVG 和图像到 SVG 生成方面取得了显著进展，但难以泛化到长序列和复杂 SVG 内容，因此很难保证生成质量。 此外，这些方法大多忽视了 SVG 理解与编辑相关任务。

表1：SVG 数据集与基准比较。
NameUnderstandingEditingGenerationMulti-Dom#Samples
Datasets
ColorSVG-100K✗✗✓✗100k
SVG-Stack✗✗✓✓2.2M
MMSVG✗✗✓✓2.0M
SVGX✓✗✓✗1.0M
UniSVG✓✗✓✗525k
DeepSVG✗✗✓✓100k
SAgoge (Ours)✓✓✓✓16M
Benchmarks
SGP-Bench✓✗✗✗4.3k
SVGEditBench✗✓✗✗1.3k
VGBench✓✗✓✗10.1k
SVGenius✓✓✓✗2.3k
SArena (Ours)✓✓✓✓31k

表1：SVG 数据集与基准比较。

To overcome these limitations, we propose the InternSVG family, an integrated data–benchmark–model suite for unified SVG modeling. Specifically, we introduce SAgoge, a unified multimodal dataset that jointly supports SVG understanding, editing, and generation, covering both static graphics and dynamic animations. SAgoge combines Internet-sourced and synthetic SVG data across icons, illustrations, chemical structural formulas, and animations, totaling about 16 million training samples. Comparison of SAgoge and other datasets is shown in Table 1.

为克服上述局限，我们提出 InternSVG 系列，一个集数据、基准和模型于一体、面向统一 SVG 建模的体系。 具体而言，我们推出 SAgoge，一个同时支持 SVG 理解、编辑和生成的统一多模态数据集，涵盖静态图形与动态动画。 SAgoge 融合了来自互联网和合成的 SVG 数据，覆盖图标、插图、化学结构式与动画，总计约 1,600 万个训练样本。 SAgoge 与其他数据集的比较见 表 1。

To the best of our knowledge, SAgoge is the largest multimodal SVG dataset to date, with the broadest task coverage and the most comprehensive range of difficulty. It includes two understanding tasks, ten editing tasks, and four generation tasks, covering semantic understanding, icon editing and generation, long-sequence illustration generation, animation generation, and the generation of scientific diagrams. To enable rigorous and comparable assessment, we propose SArena, a companion benchmark that standardizes tasks and metrics for understanding, editing, and generation, with sub-benchmarks covering icons, illustrations, chemical structural formulas, and animations. Together, SAgoge and SArena provide the scale, diversity, full task coverage, and standardized measurement needed to move beyond fragmented evaluations and enable systematic investigation of general SVG modeling in a unified setting.

据我们所知，SAgoge 是迄今规模最大的多模态 SVG 数据集，任务覆盖最广，难度范围也最全面。 它包含两项理解任务、十项编辑任务和四项生成任务，涵盖语义理解、图标编辑与生成、长序列插图生成、动画生成以及科学图示生成。 为实现严格且可比较的评估，我们提出配套基准 SArena，统一理解、编辑和生成任务及其指标，子基准覆盖图标、插图、化学结构式和动画。 SAgoge 与 SArena 共同提供了规模、多样性、完整的任务覆盖及标准化衡量方式，从而突破零散评估的局限，支持在统一设置下系统研究通用 SVG 建模。

Building on SAgoge and the accompanying SArena for systematic evaluation, we introduce InternSVG, a unified MLLM for SVG understanding, editing, and generation. InternSVG augments a pretrained vision–language backbone with SVG-specific tokenization, introducing compact special tokens for tags, attributes, and coordinates to reduce sequence length while retaining geometric and hierarchical structure. These tokens are initialized with a subword-based strategy that anchors them in the pretrained embedding space, stabilizing early training and accelerating convergence. Training adopts a two-stage strategy that progresses from short static SVGs to longer illustrations and complex animations.

以 SAgoge 为基础，并使用配套的 SArena 进行系统评估，我们推出 InternSVG，一个面向 SVG 理解、编辑和生成的统一 MLLM。 InternSVG 在预训练视觉语言主干上增加 SVG 专用分词机制，为标签、属性和坐标引入紧凑的特殊 token，在保留几何与层级结构的同时缩短序列。 这些 token 采用基于子词的策略初始化，将其置于预训练嵌入空间中，从而稳定早期训练并加快收敛。 训练采用两阶段策略，从短序列静态 SVG 逐步过渡到更长的插图和复杂动画。

Through extensive experiments, we demonstrate that unified modeling can effectively improve performance across understanding, editing, and generation tasks. Comprehensive evaluations further show that our InternSVG surpasses both open-source and proprietary models on SArena and previous benchmarks. For example, on the SArena-Icon benchmark, InternSVG surpasses Claude-Sonnet-4, the strongest proprietary baseline on SVG tasks, by about 11% higher acc in understanding tasks, 34% higher PSNR in editing tasks, 56% lower FID in Text-to-SVG tasks, and 22% higher SSIM in Image-to-SVG tasks.

通过大量实验，我们证明统一建模能够有效提升理解、编辑和生成任务的性能。 全面评估进一步表明，InternSVG 在 SArena 和已有基准上均超越了开源模型与专有模型。 例如，在 SArena-Icon 基准上，相比 SVG 任务中最强的专有基线 Claude-Sonnet-4，InternSVG 在理解任务上的准确率约高 11%，在编辑任务上的 PSNR 约高 34%，在文本到 SVG 任务上的 FID 约低 56%，在图像到 SVG 任务上的 SSIM 约高 22%。

In summary, our contributions are below:

综上，我们的贡献如下：

  1. We construct SAgoge, the largest and most comprehensive multimodal SVG dataset to date, encompassing static graphics and animations with over 16 million training samples. To enable rigorous and comparable evaluation, we further establish SArena, a companion benchmark that standardizes tasks and metrics across SVG understanding, editing, and generation.
  1. We propose InternSVG, a unified MLLM for SVG understanding, editing, and generation. It introduces SVG-specific tokenization with subword-initialized special tokens and adopts a two-stage training strategy to support effective cross-task generalization.
  1. We conduct extensive experiments to demonstrate the benefits of unified modeling. The results on SArena and prior benchmarks show that our InternSVG outperforms traditional approaches as well as general-purpose open-source and proprietary models.
  1. 我们构建了 SAgoge，这是迄今规模最大、覆盖最全面的多模态 SVG 数据集，包含静态图形和动画，训练样本超过 1,600 万个。为实现严格且可比较的评估，我们进一步建立配套基准 SArena，统一 SVG 理解、编辑和生成任务及其指标。
  1. 我们提出 InternSVG，一个用于 SVG 理解、编辑和生成的统一 MLLM。它引入 SVG 专用分词机制，通过子词初始化特殊 token，并采用两阶段训练策略，支持有效的跨任务泛化。
  1. 我们开展了大量实验，验证统一建模的优势。在 SArena 和已有基准上的结果表明，InternSVG 优于传统方法以及通用的开源模型和专有模型。

2. Related Works ​

2.1. SVG Datasets and Benchmarks ​

Most existing SVG datasets and benchmarks are limited in task coverage or data type and remain too small for effective model training, leading to fragmented evaluations and limited insights into generalization across tasks and complexity. SGP-Bench evaluates semantic comprehension and consistency in symbolic graphics programs. SVGEditBench and its extension V2 focus narrowly on instruction-based SVG editing measured by low-level syntactic metrics. On the generative side, SVG-Stack, SVGX, MMSVG, and ColorSVG-100K address Text-to-SVG and Image-to-SVG generation, while VGBench and UniSVG jointly evaluate understanding and generation.

现有 SVG 数据集和基准大多在任务覆盖范围或数据类型方面受到限制，且规模仍不足以有效训练模型，导致评估彼此割裂，也难以充分认识模型在不同任务与复杂度之间的泛化能力。 SGP-Bench 评估模型对符号图形程序的语义理解及其一致性。 SVGEditBench 及其扩展版本 V2 聚焦于基于指令的 SVG 编辑，并使用底层语法指标进行衡量。 在生成方面，SVG-Stack、SVGX、MMSVG 和 ColorSVG-100K 涵盖文本到 SVG 与图像到 SVG 生成，而 VGBench 和 UniSVG 同时评估理解与生成。

DeepSVG introduces a dataset of 100K SVG icons and explores generation, interpolation, and latent-space animation of static and limited animated graphics, but lacks rich editing instructions and image-conditioned generation. SVGenius introduces a comprehensive benchmark covering understanding, editing, and generation with systematic complexity levels and multi-dimensional metrics, but it includes only about 2,400 queries, making it sufficient for evaluation yet inadequate for training. In contrast, our SAgoge is substantially larger and more diverse, encompassing both static graphics and SVG animations. It unifies SVG understanding, editing, and generation, and with approximately 16M task samples, its scale and diversity enable robust model training and comprehensive evaluation across the full spectrum of SVG tasks, which effectively address the coverage and scalability limitations of prior datasets.

DeepSVG 引入了包含 10 万个 SVG 图标的数据集，探索静态图形及有限动画图形的生成、插值和潜空间动画，但缺少丰富的编辑指令与以图像为条件的生成。 SVGenius 提出了涵盖理解、编辑和生成的综合基准，系统地划分复杂度等级并采用多维度指标，但仅包含约 2,400 条查询，足以用于评估，却不足以用于训练。 相比之下，我们的 SAgoge 规模更大、多样性更高，同时涵盖静态图形和 SVG 动画。 它统一了 SVG 理解、编辑和生成，包含约 1,600 万个任务样本；其规模和多样性支持稳健的模型训练及覆盖全部 SVG 任务的综合评估，有效解决了以往数据集覆盖范围有限和扩展性不足的问题。

2.2. SVG Modeling Methods ​

Early research on SVG modeling treated vector graphics as sequences of geometric primitives and relied on specialized generative architectures trained on limited domains. Approaches such as DeepSVG used hierarchical VAEs with Transformer decoders to generate icons, while optimization-based methods like DiffVG and LIVE applied differentiable rasterization to align SVG primitives with target images iteratively. Although these techniques captured low-level structure, they lacked semantic understanding, struggled with complex compositions, and were computationally costly, often producing redundant or unsimplified paths. More recent works have explored diffusion-based pipelines, such as VectorFusion and SVGDreamer, which refine vector renderings via Text-to-Image diffusion combined with SVG constraints.

早期 SVG 建模研究将矢量图形视为几何图元序列，并依赖在有限领域中训练的专用生成架构。 DeepSVG 等方法使用带有 Transformer 解码器的层次化 VAE 生成图标，而 DiffVG 和 LIVE 等基于优化的方法则采用可微栅格化，将 SVG 图元与目标图像进行迭代对齐。 尽管这些技术能够捕捉底层结构，但缺乏语义理解，难以处理复杂构图，计算成本较高，且常常产生冗余或未经简化的路径。 较新的研究探索了基于扩散模型的流程，例如 VectorFusion 和 SVGDreamer，将文本到图像扩散与 SVG 约束相结合，以优化矢量渲染结果。

These approaches improve visual fidelity but remain limited to generation and offer only restricted controllability and task diversity. The emergence of large language models (LLMs) has shifted SVG modeling toward code-centric paradigms that leverage the textual nature of SVG. Methods such as StarVector, OmniSVG, LLM4SVG, and SVGBuilder integrate visual encoders with text decoders to generate SVGs from textual or visual inputs, achieving notable advances in Text-to-SVG and Image-to-SVG synthesis. However, these approaches are fragmented across isolated tasks, with limited mechanisms to handle the increasing complexity of SVG content. Different from prior approaches, InternSVG is a unified MLLM for SVG understanding, editing, and generation. It is capable of understanding both the semantic and structural information of SVG code, editing graphics based on user instructions, and generating SVGs from textual or visual prompts.

这些方法提高了视觉保真度，但仍局限于生成任务，可控性和任务多样性均较为有限。 大语言模型（LLM）的出现推动 SVG 建模转向以代码为中心的范式，利用 SVG 本身的文本特性。 StarVector、OmniSVG、LLM4SVG 和 SVGBuilder 等方法将视觉编码器与文本解码器结合，根据文本或视觉输入生成 SVG，在文本到 SVG 和图像到 SVG 合成方面取得了显著进展。 然而，这些方法分散于彼此孤立的任务，处理日益复杂的 SVG 内容的机制仍然有限。 不同于以往方法，InternSVG 是一个面向 SVG 理解、编辑和生成的统一 MLLM。 它既能理解 SVG 代码的语义与结构信息，也能根据用户指令编辑图形，并根据文本或视觉提示生成 SVG。

3. Dataset and Benchmark ​

数据集构建流程概览。 原始 SVG 来自网络收集和专门构建的合成流程，随后被规范化到 128\times128 的画布并简化以缩短代码。渲染后的图像或视频、处理后的 SVG 代码以及人工编写的提示词被送入 MLLM，以合成用于理解、编辑和生成的高质量训练样本。

图2：数据集构建流程概览。 原始 SVG 来自网络收集和专门构建的合成流程，随后被规范化到 128×128 的画布并简化以缩短代码。渲染后的图像或视频、处理后的 SVG 代码以及人工编写的提示词被送入 MLLM，以合成用于理解、编辑和生成的高质量训练样本。

图2：数据集构建流程概览。 原始 SVG 来自网络收集和专门构建的合成流程，随后被规范化到 128×128 的画布并简化以缩短代码。渲染后的图像或视频、处理后的 SVG 代码以及人工编写的提示词被送入 MLLM，以合成用于理解、编辑和生成的高质量训练样本。

3.1. SAgoge Dataset ​

We introduce SAgoge, a large-scale and comprehensive dataset for SVG tasks with more than 16 million training samples spanning icons, illustrations, chemical structures, and animations. The construction pipeline is illustrated in Figure 2.

我们推出 SAgoge，一个规模庞大、覆盖全面的 SVG 任务数据集，包含超过 1,600 万个训练样本，涵盖图标、插图、化学结构和动画。 其构建流程如 图 2 所示。

3.1.1. Task Definition ​

SAgoge is a comprehensive resource that supports unified modeling of SVG understanding, editing, and generation. We define a fine-grained task suite spanning diverse vector-graphic categories to enable consistent training and evaluation. Understanding comprises SVG description and multiple-choice question answering. Editing for icons contains ten subtasks organized by difficulty, including eight low-level operations (color editing, stroke addition, translation, scaling, rotation, flipping, transparency adjustment, and cropping) and two high-level tasks (semantic color editing and style transfer). Generation encompasses Text-to-SVG and Image-to-SVG for icons, illustrations, and chemical structures, as well as Text-to-SANI and Video-to-SANI for animations. Appendix C.3.3 presents examples of tasks in SAgoge.

SAgoge 是支持 SVG 理解、编辑和生成统一建模的综合资源。 我们定义了覆盖多种矢量图形类别的细粒度任务体系，以支持一致的训练与评估。 理解任务包括 SVG 描述和选择题问答。 图标编辑包含按难度组织的十项子任务，其中包括八项低层级操作（颜色编辑、添加描边、平移、缩放、旋转、翻转、透明度调整和裁剪）以及两项高层级任务（语义颜色编辑和风格迁移）。 生成任务涵盖面向图标、插图和化学结构的文本到 SVG 与图像到 SVG，以及面向动画的文本到 SANI 与视频到 SANI。 附录 C.3.3 展示了 SAgoge 中的任务示例。

3.1.2. Dataset Creation Pipeline ​

Data Source. SAgoge integrates Internet-sourced and synthetic SVGs. Icons and a subset of illustrations and animations are collected from public repositories such as Iconfont, SVGRepo, and OpenClipart. Chemical structures are generated by converting PubChem SDF files to SVG with the Open Babel toolkit. Due to the scarcity of open-source SVGs for illustrations and animations, we build a dedicated data synthesis pipeline to expand coverage in these domains. Implementation details are provided in Appendix C.3 and data statistics of SAgoge are summarized in Table 24.

数据来源。 SAgoge 整合了来自互联网的 SVG 与合成 SVG。 图标以及部分插图和动画从 Iconfont、SVGRepo 和 OpenClipart 等公共资源库收集。 化学结构通过 Open Babel 工具包将 PubChem 的 SDF 文件转换为 SVG 而生成。 由于开源 SVG 插图和动画较为稀缺，我们构建了专门的数据合成流程，以扩大这些领域的覆盖范围。 实现细节见 附录 C.3，SAgoge 的数据统计汇总于 表 24。

Normalize and Simplify. To improve training efficiency and reduce complexity, all SVGs are normalized to a 128×128 viewBox. Because collected files are often verbose and LLMs operate under limited context lengths, we further simplify the code by removing nonessential metadata, comments, and redundant declarations while preserving the core geometric primitives, semantic information, and hierarchical group structures. This normalization and simplification shorten token sequences, improve consistency, and make the corpus better suited for scalable training and evaluation.

规范化与简化。 为提高训练效率并降低复杂度，所有 SVG 均被规范化到 128×128 的 viewBox。 由于收集的文件通常较为冗长，而 LLM 的上下文长度有限，我们进一步删除不必要的元数据、注释和冗余声明以简化代码，同时保留核心几何图元、语义信息以及层次化分组结构。 这种规范化和简化缩短了 token 序列，提高了数据一致性，使语料更适合可扩展的训练与评估。

Construction of Understanding and Generation Data. We employ distinct annotation strategies tailored to different data modalities to ensure high-quality. We first render SVGs into raster images or videos, then utilize specific MLLMs with carefully designed prompts. For understanding tasks, we feed rasterized images to MLLMs such as InternVL3 (train set) or GPT-4o (test set) to obtain fine-grained geometric and semantic descriptions. For generation tasks, icons and illustrations are annotated by providing rasterized images to an MLLM for captioning. For chemical data, we directly use the IUPAC names or common chemical names retrieved from PubChem as textual instructions. For animation data, each SVG animation is converted into an MP4 video and annotated by Gemini 2.0 Flash.

理解与生成数据的构建。 我们针对不同数据模态采用不同的标注策略，以确保数据质量。 首先将 SVG 渲染为栅格图像或视频，再结合精心设计的提示词使用特定 MLLM。 对于理解任务，我们将栅格化图像输入 InternVL3（训练集）或 GPT-4o（测试集）等 MLLM，以获得细粒度的几何和语义描述。 对于生成任务，图标和插图通过将栅格化图像输入 MLLM 来生成图像描述，完成标注。 对于化学数据，我们直接使用从 PubChem 获取的 IUPAC 名称或化学通用名称作为文本指令。 对于动画数据，每个 SVG 动画均转换为 MP4 视频，并由 Gemini 2.0 Flash 进行标注。

Construction of Editing Data. The editing dataset consists of triplets (instruction, original SVG, edited SVG) generated across three categories: simple editing, color editing, and style transfer. Simple editing operations are applied via template-based SVG code modifications and verified by Qwen2.5-VL-72B. Color editing uses code-level replacement from a 147-color palette with MLLM-generated instructions. Style transfer pairs category-consistent SVGs using CLIP (similarity > 0.8) and synthesizes instructions from paired images and captions. Qwen2.5-VL-72B annotates the train set, while GPT-4o is used for the test set with human verification. Prompt templates for dataset construction are provided in Appendix C.4.1.

编辑数据的构建。 编辑数据集由（指令、原始 SVG、编辑后 SVG）三元组组成，涵盖简单编辑、颜色编辑和风格迁移三类。 简单编辑通过基于模板的 SVG 代码修改来执行，并由 Qwen2.5-VL-72B 进行验证。 颜色编辑在代码层面从包含 147 种颜色的调色板中进行替换，并由 MLLM 生成指令。 风格迁移使用 CLIP 对类别一致的 SVG 进行配对（相似度 > 0.8），再根据成对图像及其描述合成指令。 训练集由 Qwen2.5-VL-72B 标注，测试集则由 GPT-4o 标注并经人工验证。 数据集构建所用的提示词模板见 附录 C.4.1。

3.2. SArena Benchmark ​

To enable systematic evaluation across SVG understanding, editing, and generation, we introduce SArena, a benchmark that aligns with the domains and difficulty spectrum covered by SAgoge and provides standardized tasks and metrics. SArena includes 4 sub-benchmarks, i.e., icons, illustrations, chemical structures, and animation. To ensure reliability and fairness of evaluation, all SVG samples used in SArena are carefully curated through a combination of automated preprocessing and manual screening, with low-quality, corrupted, and semantically ambiguous files removed. Data statistics of SArena are shown in Appendix C.2.

为系统评估 SVG 理解、编辑和生成，我们推出 SArena，该基准与 SAgoge 覆盖的领域及难度范围相对应，并提供标准化任务与指标。 SArena 包括四个子基准，即图标、插图、化学结构和动画。 为确保评估的可靠性和公平性，SArena 使用的全部 SVG 样本均通过自动预处理与人工筛选相结合的方式精心整理，剔除了低质量、损坏及语义模糊的文件。 SArena 的数据统计见 附录 C.2。

3.2.1. Evaluation Metrics ​

Understanding. We evaluate code-level comprehension with four-option multiple-choice QA. The model receives only the SVG code and must infer the answer from its semantics and structure, and performance is reported as accuracy.

理解。 我们使用四选一问答评估代码层面的理解能力。 模型仅接收 SVG 代码，必须根据其语义与结构推断答案，性能以准确率报告。

Editing. The model edits the original SVG code to follow textual instructions. Performance is evaluated on rendered outputs using DINO score, Structural Similarity Index (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), and Peak Signal-to-Noise Ratio (PSNR).

编辑。 模型按照文本指令编辑原始 SVG 代码。 我们使用 DINO 分数、结构相似性指数（SSIM）、学习感知图像块相似度（LPIPS）和峰值信噪比（PSNR）评估渲染结果。

Text-to-SVG. The model synthesizes SVGs from textual instructions. Quality is evaluated with FID and FID-CLIP for distributional and semantic fidelity, CLIP-T2I for text–SVG alignment and CLIP-I2I for visual similarity.

文本到 SVG。 模型根据文本指令合成 SVG。 我们使用 FID 和 FID-CLIP 评估分布与语义保真度，使用 CLIP-T2I 评估文本与 SVG 的对齐程度，并使用 CLIP-I2I 评估视觉相似性。

Image-to-SVG. The model receives a reference image and textual instructions and generates the corresponding SVG. Evaluation mirrors the editing task and uses DINO score, SSIM, LPIPS, and PSNR to measure visual similarity between the rendered SVG and the ground-truth image.

图像到 SVG。 模型接收参考图像和文本指令，并生成相应 SVG。 评估方式与编辑任务一致，使用 DINO 分数、SSIM、LPIPS 和 PSNR 衡量渲染后的 SVG 与真实图像之间的视觉相似性。

Text-to-SANI. The model generates SVG animations from textual instructions. Evaluation uses FVD for distributional consistency and ViCLIP-based CLIPScores, with CLIP-T2V quantifying text–animation alignment and CLIP-V2V measuring visual similarity to reference videos.

文本到 SANI。 模型根据文本指令生成 SVG 动画。 评估采用 FVD 衡量分布一致性，并使用基于 ViCLIP 的 CLIPScore，其中 CLIP-T2V 量化文本与动画的对齐程度，CLIP-V2V 衡量其与参考视频的视觉相似性。

Video-to-SANI. In this task the model generates SVG animations conditioned on an input video and textual instructions, enabling cross-modal animation generation. Evaluation follows the Image-to-SVG protocol and uses DINO, SSIM, LPIPS, and PSNR computed between rendered SVG frames and reference frames on sampled timestamps, with scores averaged to produce the final result.

视频到 SANI。 在该任务中，模型以输入视频和文本指令为条件生成 SVG 动画，实现跨模态动画生成。 评估遵循图像到 SVG 的协议，在采样时间点计算渲染后的 SVG 帧与参考帧之间的 DINO、SSIM、LPIPS 和 PSNR，再对各帧分数取平均，得到最终结果。

In addition to the task-specific metrics, we also account for cases where the model generates syntactically incorrect or incomplete SVGs that cannot be rendered. Any unrenderable SVG is replaced with a pure black image or video as a penalty during metric computation, ensuring that the evaluation fairly reflects both the quality of valid outputs and the model’s robustness in generation.

除各任务专用指标外，我们也考虑模型生成的 SVG 存在语法错误或内容不完整、从而无法渲染的情况。 计算指标时，任何无法渲染的 SVG 都会被替换为纯黑图像或视频作为惩罚，以确保评估公平反映有效输出的质量与模型生成的稳健性。

4. Method ​

（a）InternSVG 的整体架构。（b）在分词器中加入自定义特殊 token 前后，每个 SVG 的 token 数量分布。（c）基于子词的嵌入初始化与随机初始化的训练损失曲线比较。

图3：（a）InternSVG 的整体架构。（b）在分词器中加入自定义特殊 token 前后，每个 SVG 的 token 数量分布。（c）基于子词的嵌入初始化与随机初始化的训练损失曲线比较。

图3：（a）InternSVG 的整体架构。（b）在分词器中加入自定义特殊 token 前后，每个 SVG 的 token 数量分布。（c）基于子词的嵌入初始化与随机初始化的训练损失曲线比较。

4.1. Model Architecture ​

As illustrated in Figure 3, InternSVG follows the “ViT–MLP–LLM” paradigm, using InternViT-300M as the vision encoder and Qwen2.5-7B as the language model. We further design SVG-specific special tokens and introduce a tailored embedding initialization strategy to incorporate SVG content effectively.

如 图 3 所示，InternSVG 遵循“ViT–MLP–LLM”范式，使用 InternViT-300M 作为视觉编码器，Qwen2.5-7B 作为语言模型。 我们进一步设计了 SVG 专用特殊 token，并引入有针对性的嵌入初始化策略，以有效融入 SVG 内容。

Special tokens. We design 55 tag tokens and 42 attribute tokens based on SVG grammar. Tag tokens cover core structural elements such as svg, path, and circle, as well as animation elements like animate, animateMotion, and animateTransform. Attribute tokens include common geometric attributes (e.g., viewBox, cx, cy, d) and animation-related attributes (e.g., dur, from, to, repeatCount). To represent numerical values, we add integer tokens from -128 to 128 and 100 decimal tokens from .0 to .99. This fine-grained number representation allows accurate modeling of geometry while keeping the tokenization compact. As shown in Figure 3(b), these special tokens substantially reduce sequence length compared with the original tokenizer, improving representation efficiency and alleviating computational burden.

特殊 token。 我们依据 SVG 语法设计了 55 个标签 token 和 42 个属性 token。 标签 token 涵盖 svg、path 和 circle 等核心结构元素，以及 animate、animateMotion 和 animateTransform 等动画元素。 属性 token 包括常见几何属性（例如 viewBox、cx、cy、d）和动画相关属性（例如 dur、from、to、repeatCount）。 为表示数值，我们添加了从 -128 到 128 的整数 token，以及从 .0 到 .99 的 100 个小数 token。 这种细粒度的数值表示能在保持分词紧凑的同时，对几何信息进行精确建模。 如 图 3(b) 所示，与原始分词器相比，这些特殊 token 显著缩短了序列长度，提高了表示效率，并减轻了计算负担。

Embedding initialization. Instead of randomly initializing the SVG-specific special tokens, we adopt a subword-based strategy to ensure semantic coherence and stable training. Each token is decomposed into subwords using the pretrained tokenizer, and the embeddings of these subwords are averaged to form the token embedding. Formally, let a new special token tnew be decomposed into n subwords {s1,s2,…,sn}, with corresponding embeddings {es1,es2,…,esn}. The initialized embedding etnew is computed as:

嵌入初始化。 对于 SVG 专用特殊 token，我们不采用随机初始化，而是采用基于子词的策略，以保证语义连贯性与训练稳定性。 每个 token 先由预训练分词器拆分为子词，再对这些子词的嵌入取平均，得到该 token 的嵌入。 形式化地，设一个新的特殊 token tnew 被拆分为 n 个子词 {s1,s2,…,sn}，对应的嵌入为 {es1,es2,…,esn}。 初始化后的嵌入 etnew 计算如下：

(1)etnew=1n∑i=1nesi

This strategy preserves the semantic prior from the original vocabulary and accelerates adaptation to SVG tokens. As shown in Figure 3(c), it reduces initial loss and accelerates convergence, demonstrating its effectiveness in stabilizing early training and enhancing overall efficiency.

这一策略保留了原始词表中的语义先验，并加快了对 SVG token 的适应。 如 图 3(c) 所示，它降低了初始损失并加速收敛，表明该策略能够稳定训练早期过程并提高整体效率。

4.2. Two-stage training strategy ​

The design of our two-stage training strategy is motivated by the inherent imbalance in SVG corpora. Icons are easy to collect in large quantities and have simpler structures, while illustrations are more difficult to obtain, less abundant, and usually longer and more complex. This imbalance in scale and complexity makes unified training challenging. To address this issue, we adopt a curriculum-style approach that progresses from simple to complex samples.

两阶段训练策略的设计源于 SVG 语料固有的不平衡性。 图标易于大量收集，结构也较为简单；插图则更难获取、数量更少，且通常更长、更复杂。 这种规模和复杂度上的不平衡使统一训练面临挑战。 为解决这一问题，我们采用类似课程学习的方法，从简单样本逐步过渡到复杂样本。

Initially, training is conducted only on the Icon and Chemistry datasets, which contain shorter and simpler SVGs. This stage covers SVG description, all editing tasks, and two generative tasks, namely Text-to-SVG and Image-to-SVG, and allows the model to build basic representation and generation abilities. Once the model has reached preliminary convergence, training is expanded to all datasets and tasks, including longer and more diverse SVGs. To ensure balanced learning, the Icon and Chemistry data are resampled to match the proportions of the other datasets. This progressive strategy enables the model to acquire SVG knowledge gradually and stably while reducing imbalance caused by heterogeneous dataset scales and complexities.

最初仅使用 Icon 和 Chemistry 数据集进行训练，其中的 SVG 较短且较为简单。 这一阶段涵盖 SVG 描述、全部编辑任务以及文本到 SVG 和图像到 SVG 两项生成任务，使模型建立基本的表示与生成能力。 模型初步收敛后，训练扩展至所有数据集与任务，纳入更长、更多样的 SVG。 为保证均衡学习，我们对 Icon 和 Chemistry 数据重采样，使其比例与其他数据集相匹配。 这种渐进策略使模型能够逐步、稳定地获取 SVG 知识，同时减轻数据集规模与复杂度差异造成的不平衡。

5. Experiments ​

5.1. Experimental Setup ​

InternSVG builds on a pretrained InternVL3-8B model and adopts a two-stage training strategy. In stage one, we train on the Icon and Chemistry datasets that contain shorter and simpler SVGs. Icon provides 1.0M samples for understanding, 1.0M for editing, and 5.6M for generation, and Chemistry provides 2.0M generation samples. In stage two, we expand to all domains and tasks, using 350K understanding, 500K editing, and 1.1M generation samples from Icon, 1.0M generation samples from Illustration, 1.0M from Chemistry, and 120K from Animation. Optimization uses AdamW with a learning rate of 2×10−4 in stage one and 1×10−4 in stage two, running on 96 NVIDIA A800 GPUs with a per-device batch size of 1.

InternSVG 基于预训练的 InternVL3-8B 模型构建，并采用两阶段训练策略。 第一阶段使用包含较短、较简单 SVG 的 Icon 和 Chemistry 数据集进行训练。 Icon 提供 100 万个理解样本、100 万个编辑样本和 560 万个生成样本，Chemistry 提供 200 万个生成样本。 第二阶段扩展至全部领域和任务，使用 Icon 中的 35 万个理解样本、50 万个编辑样本和 110 万个生成样本，以及 Illustration 中的 100 万个生成样本、Chemistry 中的 100 万个生成样本和 Animation 中的 12 万个生成样本。 优化采用 AdamW，第一阶段学习率为 2×10−4，第二阶段为 1×10−4，使用 96 张 NVIDIA A800 GPU 训练，每张设备的批大小为 1。

5.2. Quantitative Evaluations ​

To validate the effectiveness of our SAgoge dataset and InternSVG, we evaluate multiple models on the SArena benchmark. The evaluation covers open-source MLLMs, proprietary systems, traditional SVG generation methods, and LLM-based approaches, including LLM4SVG, StarVector, and OmniSVG. The complete list of evaluated models with their parameter scales is given in Appendix B.1.

为验证 SAgoge 数据集和 InternSVG 的有效性，我们在 SArena 基准上评估了多种模型。 评估涵盖开源 MLLM、专有系统、传统 SVG 生成方法，以及 LLM4SVG、StarVector 和 OmniSVG 等基于 LLM 的方法。 所评估模型及其参数规模的完整列表见 附录 B.1。

表2：SArena-Icon 上的 SVG 理解、编辑与生成比较。用大写粗体首字母表示各子指标：Overall（总体）、Color（颜色）、Geometry（几何）、Quantity（数量）和 Semantics（语义）。
ModelUnderstandingEditingText-to-SVGImage-to-SVG
OCGQSDINO ↑SSIM ↑LPIPS ↓PSNR ↑FID ↓FID-C ↓CLIP-T2I ↑CLIP-I2I ↑TokensDINO ↑SSIM ↑LPIPS ↓PSNR ↑Tokens
Traditional SVG methods
IconShop–––––––––32.28817.91920.89470.9221.5k–––––
VectorFusion–––––––––16.59417.39422.99270.30833k–––––
SVGDreamer–––––––––26.61233.31220.32969.975132k–––––
DiffVG––––––––––––––0.8690.9270.09723.61417k
LIVE––––––––––––––0.9730.9860.02435.41918k
VTracer––––––––––––––0.9660.8750.05421.7484.4k
Large language models
Qwen2.5-VL-7B52.869.350.434.956.40.9090.7280.19225.40224.78115.45421.53871.3842490.7810.5060.3786.534281
InternVL3-8B59.579.159.338.261.30.9210.7610.17029.61523.06114.30321.89771.4502690.8120.5570.3617.220256
Llama-4-Maverick64.787.562.047.262.30.9660.8700.10946.94414.9316.52623.57075.8162650.8630.5960.3298.027255
Qwen2.5-VL-72B63.482.465.144.661.60.9610.8490.12441.00615.9489.87522.94673.6812750.8370.5840.3467.834372
InternVL3-78B65.386.471.048.854.90.9580.8480.11640.53317.58010.59622.80573.1232520.8500.5840.3397.802234
GPT-4o71.088.278.547.569.60.9680.8870.08855.25515.1786.76324.61777.7422460.8740.6160.3168.435231
Gemini-2.5-Flash73.090.181.953.067.20.9420.8150.11354.20016.7205.20824.65878.2184510.8760.5870.3168.324533
Claude-Sonnet-477.191.582.453.880.60.9790.9150.07157.59515.8404.29125.42180.5794440.9150.6650.2769.855541
StarVector 8B––––––––––––––0.8710.6230.20613.595951
LLM4SVG 7B–––––––––21.9398.61119.45870.7267050.7480.4720.4095.375485
OmniSVG 3B–––––––––28.29211.31821.67974.8311.7k0.8940.7560.18612.6692.4k
InternSVG 8B85.193.085.861.999.70.9890.9520.03677.3318.7151.87623.91680.9111.0k0.9490.8110.12718.2261.3k
+8.0+1.5+2.8+8.1+19.1+0.010+0.037+0.035+19.736+6.216+2.415-1.505+0.332+0.034+0.055+0.059+4.631

表2：SArena-Icon 上的 SVG 理解、编辑与生成比较。用大写粗体首字母表示各子指标：Overall（总体）、Color（颜色）、Geometry（几何）、Quantity（数量）和 Semantics（语义）。

表3：SArena-Illustration 上的 SVG 生成结果比较。
ModelText-to-SVGImage-to-SVG
FID ↓FID-C ↓CLIP-T2I ↑CLIP-I2I ↑TokensDINO ↑SSIM ↑LPIPS ↓PSNR ↑Tokens
Traditional SVG methods
VectorFusion28.19815.66122.63871.22033k–––––
SVGDreamer32.41023.42721.02669.775132k–––––
DiffVG–––––0.8700.8860.13821.60517k
LIVE–––––0.9630.9480.07829.05518k
VTracer–––––0.9650.8790.09321.75410k
Large language models
Qwen2.5-VL-7B37.90328.45518.06961.9287560.7390.5130.4137.7321.2k
InternVL3-8B36.73625.68218.49361.9644930.7720.5690.3978.542716
Llama-4-Maverick30.83514.83121.87267.3665510.8390.6440.34010.469608
Qwen2.5-VL-72B29.52118.40720.92365.3495270.8080.6280.3639.900886
InternVL3-78B30.45719.19520.57764.8264540.8300.6380.3489.985514
GPT-4o28.12414.15023.63770.6964730.8500.6630.32710.723484
Gemini-2.5-Flash28.8658.89424.80074.7961.2k0.8290.5160.3599.0911.8k
Claude-Sonnet-427.2947.64023.09474.5251.0k0.9010.6700.30511.7311.3k
LLM4SVG 7B48.70429.56815.46862.9331.2k0.7130.4940.4136.221476
OmniSVG 3B42.75622.88516.86164.8154.5k0.7970.6560.33010.4336.7k
InternSVG 8B22.3975.14121.11674.6628.1k0.9240.7160.18814.6447.7k
+4.897+2.499-3.684-0.134+0.023+0.046+0.117+2.913

表3：SArena-Illustration 上的 SVG 生成结果比较。

表4：SArena-Chemistry 和 SArena-Animation 上的 SVG 生成结果比较。
ModelChemistry: Text-to-SVGChemistry: Image-to-SVGAnimation: Text-to-SANIAnimation: Video-to-SANI
FID ↓FID-C ↓CLIP-I2I ↑DINO ↑SSIM ↑LPIPS ↓PSNR ↑FVD ↓CLIP-T2V ↑CLIP-V2V ↑DINO ↑SSIM ↑LPIPS ↓PSNR ↑
Qwen2.5-VL-7B56.24873.69851.8140.7690.4680.2747.501214.37919.11850.6490.7870.7160.27311.758
InternVL3-8B33.61361.67556.8560.8650.7830.20313.840310.06617.01743.8560.7800.6120.2869.883
Llama-4-Maverick26.84431.92469.6430.9080.7980.17314.977141.47022.30467.6150.8410.7540.24612.858
Qwen2.5-VL-72B32.30744.54063.9310.8460.6470.21512.106151.68220.37659.4540.8340.7210.26111.931
InternVL3-78B29.21640.08065.9690.9110.8130.17715.375169.15920.26360.8960.8280.7040.26411.336
GPT-4o24.50519.29776.5990.9200.7910.17414.673155.39322.80870.6080.8600.7430.25012.260
Gemini-2.5-Flash27.70821.77775.8970.9340.8170.15515.539151.98322.23966.5540.8470.7010.25712.015
Claude-Sonnet-421.25215.24078.3080.9570.8710.13217.554169.48424.07074.1790.8670.7600.24013.189
StarVector 8B–––0.9770.8410.14717.419–––––––
InternSVG 8B9.9740.87793.9310.9940.8730.13817.72299.47422.57273.1620.8760.7540.23714.168
+11.278+14.363+15.623+0.017+0.002-0.006+0.178+41.996-1.498-1.017+0.009-0.006+0.003+0.979

表4：SArena-Chemistry 和 SArena-Animation 上的 SVG 生成结果比较。

As shown in Table 2, we systematically evaluate various models on SArena-Icon across understanding, editing, and generation tasks. Our InternSVG outperforms the second-best model by 8 points in Overall accuracy. In editing, InternSVG achieves the best scores across DINO, SSIM, LPIPS, and PSNR, yielding higher visual fidelity and geometric consistency. For Text-to-SVG, it achieves the best FID, FID-C, and CLIP-I2I among all methods, while its CLIP-T2I is comparable to that of Claude-Sonnet-4. In Image-to-SVG, optimization-based methods remain strong on similarity, yet InternSVG achieves comparable visual similarity while producing much more compact code at about 1.3k tokens per SVG, roughly one fourteenth of LIVE’s token count, and it unifies understanding, editing, and generation within a single model.

如 表 2 所示，我们在 SArena-Icon 上系统评估了多种模型在理解、编辑和生成任务中的表现。 InternSVG 的总体准确率比次优模型高出 8 个百分点。 在编辑任务中，InternSVG 在 DINO、SSIM、LPIPS 和 PSNR 上均取得最佳分数，具有更高的视觉保真度和几何一致性。 对于文本到 SVG，它在所有方法中取得最佳 FID、FID-C 和 CLIP-I2I，而 CLIP-T2I 与 Claude-Sonnet-4 相当。 在图像到 SVG 中，基于优化的方法在相似性方面仍然很强，但 InternSVG 在获得相当视觉相似性的同时，生成的代码更加紧凑，每个 SVG 约含 1,300 个 token，约为 LIVE 的十四分之一，并在单个模型中统一了理解、编辑和生成。

As shown in Table 3, on SArena-Illustration, our method achieves the best FID and FID-C in Text-to-SVG, with CLIPScores comparable to Gemini-2.5-Flash, the best-performing proprietary system, and outperforms both general MLLMs and LLM-based methods in Image-to-SVG. On SArena-Chemistry, where common or IUPAC names are used as prompts, existing MLLMs perform poorly on Text-to-SVG due to limited training on scientific graphics, while InternSVG achieves the best results on both tasks with a large margin in Text-to-SVG (Table 4). On SArena-Animation, InternSVG follows a protocol of sampling 8 frames from video input and surpasses all open-source MLLMs, reaching performance close to proprietary models on both Text-to-SANI and Video-to-SANI. Complete evaluation results are provided in Appendix B.2.

如 表 3 所示，在 SArena-Illustration 上，我们的方法在文本到 SVG 任务中取得最佳 FID 和 FID-C，CLIPScore 与表现最好的专有系统 Gemini-2.5-Flash 相当，并在图像到 SVG 任务上优于通用 MLLM 和基于 LLM 的方法。 在以通用名称或 IUPAC 名称作为提示的 SArena-Chemistry 上，由于科学图形训练不足，现有 MLLM 在文本到 SVG 任务上的表现较差，而 InternSVG 在两项任务上均取得最佳结果，尤其在文本到 SVG 上大幅领先（表 4）。 在 SArena-Animation 上，InternSVG 遵循从输入视频采样 8 帧的协议，超越了所有开源 MLLM，并在文本到 SANI 和视频到 SANI 上取得接近专有模型的表现。 完整评估结果见 附录 B.2。

To fully validate the effectiveness of SAgoge and InternSVG, we further evaluated the performance of InternSVG on established benchmarks, including SGP-Bench, SVG-Stack, and UniSVG. Notably, InternSVG achieves the best score on SGP-Bench, surpassing proprietary models like Claude-Sonnet-4, and dominates UniSVG with a Final Score of 0.826. On SVG-Stack, InternSVG achieves superior Dino, LPIPS, and MSE scores with comparable SSIM, using only 1.2k tokens compared to StarVector's 3.7k–5.3k. The results demonstrate that the SAgoge dataset and InternSVG framework provide significant improvements across diverse SVG tasks, proving their efficacy extends well beyond the specific data distribution of SArena. Additional results are reported in Appendix B.3, Appendix B.4, and Appendix B.5.

为全面验证 SAgoge 和 InternSVG 的有效性，我们进一步在 SGP-Bench、SVG-Stack 和 UniSVG 等已有基准上评估了 InternSVG 的性能。 值得注意的是，InternSVG 在 SGP-Bench 上取得最佳分数，超过 Claude-Sonnet-4 等专有模型，并以 0.826 的最终得分在 UniSVG 上领先。 在 SVG-Stack 上，InternSVG 取得了更优的 Dino、LPIPS 和 MSE 分数，以及相当的 SSIM，同时仅使用 1,200 个 token，而 StarVector 需要 3,700–5,300 个。 结果表明，SAgoge 数据集和 InternSVG 框架在多种 SVG 任务上带来了显著提升，证明其有效性远超 SArena 的特定数据分布。 更多结果见 附录 B.3、附录 B.4 和 附录 B.5。

5.3. Qualitative Visualization ​

InternSVG 生成的 SVG 样本可视化。

图4：InternSVG 生成的 SVG 样本可视化。

图4：InternSVG 生成的 SVG 样本可视化。

Figure 4 presents qualitative examples of SVGs generated by InternSVG. The results show that our InternSVG can produce diverse and visually appealing graphics with clear structures and semantically rich content. More qualitative visualization results of InternSVG and comparisons with other models can be found in Appendix B.9.

图 4 展示了 InternSVG 生成的 SVG 定性示例。 结果表明，InternSVG 能够生成多样且具有视觉吸引力的图形，其结构清晰、语义内容丰富。 更多 InternSVG 的定性可视化结果以及与其他模型的比较见 附录 B.9。

5.4. Ablation Studies ​

表5：不同任务组合下生成（G）、理解（U）与编辑（E）的消融结果。
TasksUText-to-SVGImage-to-SVGEditing
OverallFID ↓FID-C ↓CLIP-T2I ↑CLIP-I2I ↑DINO ↑SSIM ↑LPIPS ↓PSNR ↑DINO ↑SSIM ↑LPIPS ↓PSNR ↑
G–15.5487.84821.07373.0160.8070.5470.3876.860––––
U62.9––––––––––––
E–––––––––0.9430.8110.12542.131
G+U75.112.8816.75021.07073.3100.8130.5630.3827.042––––
G+E–13.8426.81021.09973.2790.8080.5530.3836.8970.9690.8810.09049.503
G+U+E75.412.3876.61121.19873.6300.8140.5680.3797.0950.9770.9060.07554.618

表5：不同任务组合下生成（G）、理解（U）与编辑（E）的消融结果。

Benefits of Unified SVG Modeling. To validate the effectiveness of unified SVG modeling in improving model performance, we sample 100K SVGs from the Icon dataset and construct a multi-task training set. The set includes 100K samples for multiple-choice QA and 100K samples for SVG description in the understanding task, 100K samples for semantic color editing in the editing task, and 100K samples each for text-to-SVG and image-to-SVG in the generation task. As shown in Table 5, we compare single-task, two-task, and three-task training configurations. The results show that multi-task joint training leads to consistent performance improvements and achieves the best results across the evaluation metrics of all tasks. Overall, three-task joint training yields significant gains in most metrics, which confirms that unified SVG modeling facilitates cross-task knowledge transfer, enables the learning of richer structural and semantic representations, and improves the generalization ability of the model.

统一 SVG 建模的收益。 为验证统一 SVG 建模在提升模型性能方面的有效性，我们从 Icon 数据集中采样 10 万个 SVG，并构建多任务训练集。 该训练集包含理解任务中的 10 万个选择题问答样本和 10 万个 SVG 描述样本、编辑任务中的 10 万个语义颜色编辑样本，以及生成任务中各 10 万个文本到 SVG 和图像到 SVG 样本。 如 表 5 所示，我们比较了单任务、双任务和三任务训练配置。 结果表明，多任务联合训练能够持续提升性能，并在全部任务的评估指标上取得最佳结果。 总体而言，三任务联合训练在大多数指标上取得了显著收益，证实统一 SVG 建模能够促进跨任务知识迁移，学习更丰富的结构与语义表示，并提高模型的泛化能力。

表6：两阶段训练策略效果的消融结果。
StrategyIllustration: Text-to-SVGIllustration: Image-to-SVGAnimation: Text-to-SANIAnimation: Video-to-SANI
FID ↓FID-C ↓CLIP-T2I ↑CLIP-I2I ↑DINO ↑SSIM ↑LPIPS ↓PSNR ↑FVD ↓CLIP-T2V ↑CLIP-V2V ↑DINO ↑SSIM ↑LPIPS ↓PSNR ↑
One-stage68.64425.67119.21766.4000.8300.5030.27810.345101.43322.29169.9930.8670.7350.24513.420
Two-stage22.3975.14121.11674.6620.9240.7160.18814.64499.47422.57273.1620.8760.7540.23714.168

表6：两阶段训练策略效果的消融结果。

Comparison of One-stage and Two-stage Training. To verify the effectiveness of the two-stage training strategy, we merge the data from both stages into a single-stage training scheme as a baseline. As demonstrated in Table 6, the two-stage model achieves substantial advantages on both illustration and animation generation. Especially for the long-sequenced illustration generation task, the two-stage training strategy exhibits remarkable improvements, with FID-C reduced from 25.67 to 5.14 and DINO increased from 0.830 to 0.924. More comparisons can be found in Appendix B.7.

单阶段与两阶段训练的比较。 为验证两阶段训练策略的有效性，我们将两个阶段的数据合并到单阶段训练方案中，作为基线。 如 表 6 所示，两阶段模型在插图和动画生成方面均具有明显优势。 尤其是在长序列插图生成任务上，两阶段训练策略带来了显著提升：FID-C 从 25.67 降至 5.14，DINO 从 0.830 提升至 0.924。 更多比较见 附录 B.7。

Benefits of SVG-specific special tokens and subword-based embedding initialization. To assess the contribution of these two components, we sampled 50% of the icon, illustration, and chemistry training data used in stage 2 and trained three model variants: Model Raw (no special tokens), Model T (special tokens with random initialization), and Model T+E (special tokens with subword-based initialization). As shown in Table 7, Model T+E consistently delivers the best performance across both benchmarks, demonstrating that combining special tokens with subword-based embedding initialization provides the largest overall gains. Furthermore, models that incorporate special tokens (Model T and Model T+E) exhibit substantially higher success rates on illustration generation, which involves longer SVG sequences. This trend aligns with the token-compression effect shown in Figure 3(b): introducing SVG-specific tokens reduces sequence length, stabilizes training, and ultimately improves task reliability.

SVG 专用特殊 token 与基于子词的嵌入初始化的收益。 为评估这两个组件的贡献，我们从第二阶段使用的图标、插图和化学训练数据中采样 50%，训练了三个模型变体：Model Raw（不使用特殊 token）、Model T（使用随机初始化的特殊 token）以及 Model T+E（使用基于子词初始化的特殊 token）。 如 表 7 所示，Model T+E 在两个基准上始终取得最佳性能，表明将特殊 token 与基于子词的嵌入初始化相结合能够带来最大的整体收益。 此外，在涉及更长 SVG 序列的插图生成任务中，引入特殊 token 的模型（Model T 和 Model T+E）的成功率显著更高。 这一趋势与 图 3(b) 中展示的 token 压缩效果一致：引入 SVG 专用 token 能够缩短序列长度、稳定训练，最终提高任务可靠性。

表7：SVG 专用特殊 token 与基于子词的嵌入初始化效果的消融实验。SR 表示成功率。
ModelUText-to-SVGImage-to-SVG
OverallSRFID ↓FID-C ↓CLIP-T2I ↑CLIP-I2I ↑SRDINO ↑SSIM ↑LPIPS ↓PSNR ↑
SArena-Icon
Raw79.392.90%15.7885.11222.76776.88484.30%0.8090.4930.3816.571
T79.797.62%11.9224.57823.27577.37489.59%0.9190.7410.16614.342
T + E80.898.42%11.5994.50023.34877.54195.64%0.9370.7950.13915.837
SArena-Illustration
Raw–57.22%68.30126.10818.87066.22250.08%0.7350.3310.4126.053
T–69.02%53.08022.35819.17466.46957.37%0.8070.4480.3079.424
T + E–78.81%42.81717.43819.19467.37975.31%0.8620.5780.25711.810

表7：SVG 专用特殊 token 与基于子词的嵌入初始化效果的消融实验。SR 表示成功率。

6. Conclusions ​

In this work, we introduce the InternSVG family, a unified data-benchmark-model suite for scalable vector graphics. It integrates the large-scale and diverse SAgoge, the comprehensive benchmark SArena, and the unified MLLM InternSVG. SAgoge encompasses a broad spectrum of SVG tasks with varied difficulty levels, ranging from static icons and scientific diagrams to long-sequence illustrations and dynamic animations. SArena provides a standardized evaluation framework that assesses both mainstream MLLMs and traditional SVG generation approaches, enabling consistent comparison across tasks and domains. Through extensive experiments, we demonstrate that this unified formulation enhances SVG understanding, editing, and generation, yielding consistent improvements across domains and tasks. Our study highlights the importance of unified SVG modeling, and we hope it can serve as an insight for future research on vector-graphic reasoning and multimodal intelligence.

本文提出 InternSVG 系列，一个面向可缩放矢量图形、集数据、基准和模型于一体的统一体系。 它整合了规模庞大且多样的 SAgoge、综合基准 SArena，以及统一 MLLM InternSVG。 SAgoge 涵盖难度各异的广泛 SVG 任务，从静态图标和科学图示，到长序列插图与动态动画。 SArena 提供了标准化评估框架，同时评估主流 MLLM 和传统 SVG 生成方法，支持跨任务、跨领域的一致比较。 通过大量实验，我们证明这种统一建模方式能够增强 SVG 理解、编辑和生成，并在各领域与任务中带来持续提升。 本研究强调了统一 SVG 建模的重要性，我们希望它能为未来矢量图形推理与多模态智能研究提供启发。