---

Compositional SVG Generation via VLM-Driven Hierarchical Semantic Parsing

SVGVectorizationVLMEMNLP 2026高丽大学KAIST

Park S, Kim T, Han G, et al. Compositional SVG generation via VLM-driven hierarchical semantic parsing[PP/OL]. V1. arXiv (2026-09-13)[2026-10-23]. DOI:10.48550/arXiv.2609.14657.

https://arxiv.org/abs/2609.14657v1

---

基于 VLM 驱动的层级语义解析实现组合式 SVG 生成

Abstract

While Vision-Language Models (VLMs) excel at visual reasoning, generating structured, editable Scalable Vector Graphics (SVG) remains a fundamental challenge. Existing pipelines predominantly yield flat, semantically agnostic collections of paths, where editing a single object requires manually identifying its constituent paths. To address this, we propose a VLM-driven agentic framework for semantic compositional SVG generation. Our pipeline recursively parses visual scenes into semantic and geometric hierarchies via top-down decomposition, visual grounding, and prompt-driven amodal occlusion recovery, ensuring each component is geometrically complete. Furthermore, we introduce the Semantic SVG Benchmark with human-annotated semantic groups and novel sub-component metrics (Semantic Recall/Precision, PERE) to explicitly evaluate structural compositionality and functional editability. Experiments show that our natively predicted structures surpass the upper bounds of existing flat-generation methods in both grouping quality and editability, while maintaining state-of-the-art visual fidelity.

尽管视觉语言模型（VLM）擅长视觉推理，但生成结构化、可编辑的可缩放矢量图形（SVG）仍然是一项根本性挑战。 现有流程主要生成扁平、缺乏语义信息的路径集合，编辑其中的单个对象时，需要手动识别构成该对象的路径。 为解决这一问题，我们提出了一种由 VLM 驱动的智能体框架，用于生成具备语义组合结构的 SVG。 我们的流程通过自顶向下的分解、视觉定位以及提示驱动的全形（amodal）遮挡恢复，将视觉场景递归解析为语义与几何层级，确保每个组件在几何上都是完整的。 此外，我们提出了带有人工标注语义分组的 Semantic SVG Benchmark，并引入新的子组件指标（语义召回率／精确率、PERE），以明确评估结构组合性与功能性可编辑性。 实验表明，我们原生预测的结构在分组质量和可编辑性方面均超过了现有扁平生成方法的上界，同时保持了最先进的视觉保真度。

1. Introduction

Vision-Language Models (VLMs) have demonstrated remarkable capabilities in bridging textual semantics with visual concepts. However, translating this high-level semantic understanding into structured, manipulable visual representations remains a fundamental challenge. Scalable Vector Graphics (SVG) serve as an ideal medium to bridge this gap, as they compose images using editable mathematical primitives. Yet, true manipulability requires these primitives to be grouped into meaningful semantic structures (e.g., organizing hundreds of disparate curves into a distinct "head" or "body" group). Without such grouping, editing a single object (e.g., moving, deleting) requires manually identifying and selecting its constituent paths, whereas encoding the semantic identity and geometric completeness of each object directly in the representation makes such edits far more reliable and efficient.

视觉语言模型（VLM）在连接文本语义与视觉概念方面展现出了卓越的能力。 然而，将这种高层语义理解转化为结构化、可操控的视觉表示，仍然是一项根本性挑战。 可缩放矢量图形（SVG）使用可编辑的数学图元来构成图像，因此是弥合这一差距的理想媒介。 不过，要实现真正的可操控性，就需要将这些图元组织成有意义的语义结构（例如，将数百条分散的曲线组织成独立的“头部”或“身体”分组）。 如果没有这种分组，编辑单个对象（例如移动或删除）就需要手动识别并选择其组成路径；而将每个对象的语义身份与几何完整性直接编码到表示中，能够使此类编辑更加可靠、高效。

图1：SVG 生成中语义组合性的必要性。（a）既有方法输出扁平、无组织的路径集合，而我们的框架构建与语义对齐的层级树。（b）缺少层级语义结构时，下游编辑（例如移除王冠）会失败，因为单条路径与人类可理解的概念相互脱离。相比之下，我们的显式语义分组将整个目标对象封装为一个完整节点，从而支持成功、直观的操控。

Despite the necessity of this semantic structure, existing SVG generation pipelines predominantly yield flat, semantically agnostic outputs. Whether relying on low-level optimization-based tracing or direct sequence generation from VLMs, current methods hit a fundamental ceiling: they generate flat vector collections that entirely miss the hierarchical semantic organization necessary for intuitive human manipulation and downstream reasoning. Recovering this organization post hoc from an already vectorized SVG is likewise bounded by the given paths, which may entangle multiple objects or leave occluded regions unrecoverable.

尽管这种语义结构十分必要，现有 SVG 生成流程却主要输出扁平、缺乏语义信息的结果。 无论依赖基于低层优化的描摹，还是依赖 VLM 的直接序列生成，现有方法都会触及一个根本性瓶颈：它们生成的是扁平的矢量集合，完全缺失直观的人类操控和下游推理所需的层级语义组织。 在已经矢量化的 SVG 上事后恢复这种组织结构，同样受限于既有路径；这些路径可能将多个对象纠缠在一起，或者使被遮挡区域无法恢复。

To address this structural deficiency, we propose a paradigm shift: moving from flat SVG generation to semantic compositional SVG generation. We introduce a dynamic hierarchical parsing pipeline that leverages a multi-role VLM to recursively decompose a complex visual scene into semantically atomic sub-components. By grounding semantic concepts into pixel space, recovering occluded background geometries, and assembling the vectors into a structured markup, our approach ensures each component is geometrically complete and independently manipulable. Furthermore, because existing evaluation protocols in recent works measure only holistic, whole-image visual fidelity, they are fundamentally incapable of assessing the internal semantic validity of the generated components. To solve this, we introduce the human-annotated Semantic SVG Benchmark together with Semantic Recall and Precision, which evaluate the quality of semantic grouping, and the Post-Edit Rendering Error (PERE), which measures functional editability by simulating a structural edit.

为解决这一结构性缺陷，我们提出了一种范式转变：从扁平 SVG 生成转向具备语义组合结构的 SVG 生成。 我们引入了一套动态层级解析流程，利用承担多种角色的 VLM，将复杂视觉场景递归分解为语义上不可再分的子组件。 通过将语义概念定位到像素空间、恢复被遮挡的背景几何，并将矢量组装成结构化标记，我们的方法确保每个组件都具有完整的几何形状，并且可以独立操控。 此外，近期工作的现有评测协议只衡量整幅图像的总体视觉保真度，因此从根本上无法评估所生成组件内部的语义有效性。 为解决这一问题，我们提出了人工标注的 Semantic SVG Benchmark，以及评估语义分组质量的语义召回率和精确率，并引入编辑后渲染误差（Post-Edit Rendering Error，PERE），通过模拟结构性编辑来衡量功能性可编辑性。

Our main contributions are as follows:

我们的主要贡献如下：

• Semantic Compositional SVG Task and Benchmark: We formally propose the task of generating semantically grouped SVGs, and introduce the first benchmark with human-annotated semantic groups along with metrics that evaluate semantic compositionality and functional editability rather than mere whole-image fidelity.
• Dynamic Hierarchical Parsing Pipeline: We introduce a VLM-driven framework that parses visual scenes into compositional structures, integrating semantic grounding, targeted occlusion recovery, and assembly into a grouped SVG.
• State-of-the-Art Performance: Our experiments establish a rigorous baseline for this new task, showing that our pipeline outperforms existing flat-generation methods even when they are augmented with optimal post-hoc grouping.

• 语义组合式 SVG 任务与基准：我们正式提出生成带有语义分组的 SVG 这一任务，并引入首个带有人工标注语义分组的基准及相应指标，用于评估语义组合性和功能性可编辑性，而不仅仅是整幅图像的保真度。
• 动态层级解析流程：我们提出了一种由 VLM 驱动的框架，将视觉场景解析为组合结构，集成了语义定位、有针对性的遮挡恢复，以及带分组的 SVG 组装过程。
• 最先进的性能：我们的实验为这一新任务建立了严格的基线，并表明即使为现有扁平生成方法配备最优的事后分组，我们的流程仍然优于它们。

2. Related Work

SVG Generation Existing SVG generation methods fall into two paradigms: optimization-based approaches that reconstruct paths via low-level pixel clustering, and VLM-driven methods that predict token sequences or iteratively peel occluding layers. Both produce flat representations, omitting structural metadata such as semantically nested <g> groups. Vector Prism restores such grouping post hoc, yet remains bounded by the input vectorization, unable to subdivide entangled paths or recover occluded geometry. We instead emit paths only after decomposing the scene into meaningful components. This blind spot extends to evaluation: existing benchmarks rely on whole-image similarity, lacking part-level metrics. We therefore introduce a benchmark with human-annotated semantic groups, a Semantic Recall/Precision protocol, and the Post-Edit Rendering Error (PERE) measuring functional editability.

SVG 生成 现有 SVG 生成方法可分为两种范式：基于优化的方法通过低层像素聚类重建路径，而 VLM 驱动的方法则预测 token 序列或逐层剥离遮挡层。 两者均产生 扁平 表示，缺少语义嵌套的 <g> 分组等结构元数据。 Vector Prism 能在事后恢复此类分组，但仍受输入矢量化结果的限制，无法细分相互纠缠的路径，也无法恢复被遮挡的几何形状。 我们则先将场景分解为有意义的组件，然后才生成路径。 这一盲点也延伸到了评测：现有基准依赖整幅图像的相似度，缺少部件级指标。 因此，我们引入了带有人工标注语义分组的基准、语义召回率／精确率评测协议，以及衡量功能性可编辑性的编辑后渲染误差（PERE）。

Tool-Augmented Multimodal Agents A parallel thread augments foundation models with external tools. Building on early paradigms of tool use and reasoning-action interleaving and their scaling to broader APIs and code-as-action policies, recent work has shifted toward long-horizon, multi-tool orchestration. In multimodal settings, LLaVA-Plus equips LMMs with vision-language tool repositories, and Visual Sketchpad extends tool use to self-drawn artifacts, while concept-aware tools such as SAM 3 make tool-augmented scene parsing practical. These systems use tools as auxiliary modules for recognition, reasoning, or short-form generation. In contrast, we use them within a structural image-to-vector process, where a single VLM assumes four cooperating roles—Hierarchical Decomposer, Residual Judge, Occlusion Assessor, and Amodal Spatial Estimator—and recursively invokes segmentation, inpainting, and vectorization tools whose outputs constitute the final SVG.

工具增强的多模态智能体 另一条并行的研究路线使用外部工具来增强基础模型。 在早期工具使用、推理与动作交替执行的范式，以及将这些范式扩展到更广泛的 API 和以代码为动作的策略之后，近期工作转向了长程、多工具协同调度。 在多模态场景中，LLaVA-Plus 为大型多模态模型（LMM）配备视觉语言工具库，Visual Sketchpad 将工具使用扩展到自行绘制的产物，而 SAM 3 等具备概念感知能力的工具则使工具增强的场景解析变得可行。 这些系统将工具用作识别、推理或短篇内容生成的辅助模块。 相比之下，我们将工具用于结构化的图像到矢量转换过程：同一个 VLM 承担层级分解器、残差判别器、遮挡评估器和全形空间估计器四种相互协作的角色，并递归调用分割、修复和矢量化工具，而这些工具的输出共同构成最终的 SVG。

3. Proposed Method

Our proposed framework transforms a raster image into a semantic, compositional SVG. Rather than generating flat vector sequences autoregressively, we formulate this as a multimodal semantic parsing problem, utilizing a multi-role Vision-Language Model (VLM) to recursively parse the visual scene into a semantic and geometric hierarchy.

我们提出的框架将光栅图像转换为具有语义结构的组合式 SVG。 我们将其建模为多模态语义解析问题，利用承担多种角色的视觉语言模型（VLM），将视觉场景递归解析为语义与几何层级，而不是自回归地生成扁平矢量序列。

图2：由 VLM 驱动的组合式 SVG 生成智能体系统概览。（上）给定输入图像，VLM 通过自顶向下的递归构建语义树 T：评估每个节点在语义上是否可分解，并通过分解（按从后到前的顺序生成子标签与边界框）、定位（像素掩码）以及残差判别器来扩展可分解节点；残差判别器将未分配的残差区域提升为新节点，或将其作为噪声丢弃。循环会对每个子节点递归执行，直到所有节点都不可再分。（下）随后，VLM 判断每个叶节点是否被遮挡：对于被遮挡的叶节点（例如沙滩和遮阳伞后方的海面），先通过 VLM 估计完整轮廓，再通过修复实现全形补全；未被遮挡的叶节点则保持原样。所有组件均被矢量化，并按从后到前的顺序组装成语义 SVG S，其嵌套的 g 分组标签层级与 T 对应，从而得到可直观编辑且结构完整的输出。

3.1. Problem Formulation: Semantic Compositional SVG Generation

Our objective is to translate a visual input $\mathcal{I}$ into a semantically structured vector representation $\mathcal{S}$. We formulate this task as the generative modeling of a hierarchical semantic tree $\mathcal{T}=(\mathcal{V},\mathcal{E})$.

我们的目标是将视觉输入 $\mathcal{I}$ 转换为具有语义结构的矢量表示 $\mathcal{S}$。 我们将该任务建模为对层级语义树 $\mathcal{T}=(\mathcal{V},\mathcal{E})$ 的生成式建模。

In this directed tree, vertices $\mathcal{V}$ represent discrete visual components, and edges $\mathcal{E}$ denote compositional “part-of” relationships. Each node $v_i\in \mathcal{V}$ is defined by a tuple $v_i=(l_i,{\mathcal{S}}_i)$, where $l_i$ is an explicit textual concept identifying the region. The node set $\mathcal{V}$ is partitioned into intermediate nodes and terminal leaf nodes $\mathcal{L}$. For an intermediate node, ${\mathcal{S}}_i$ serves as a logical grouping mechanism (mapped to an SVG <g> tag) that recursively encompasses its children. For a leaf node $v_k\in \mathcal{L}$, ${\mathcal{S}}_k$ represents a semantically atomic component, which may consist of one or more mathematical vector primitives to render its complete shape. The final SVG $\mathcal{S}$ is derived by traversing $\mathcal{T}$ and nesting the primitives according to z-order (background-to-foreground) constraints.

在这棵有向树中，顶点 $\mathcal{V}$ 表示离散的视觉组件，边 $\mathcal{E}$ 表示“部分属于整体”的组合关系。 每个节点 $v_i\in \mathcal{V}$ 都由元组 $v_i=(l_i,{\mathcal{S}}_i)$ 定义，其中 $l_i$ 是用于标识该区域的显式文本概念。 节点集合 $\mathcal{V}$ 分为中间节点和终端叶节点 $\mathcal{L}$。 对于中间节点，${\mathcal{S}}_i$ 充当逻辑分组机制（映射为 SVG 的 <g> 标签），递归地包含其子节点。 对于叶节点 $v_k\in \mathcal{L}$，${\mathcal{S}}_k$ 表示一个语义上不可再分的组件，它可以由一个或多个数学矢量图元构成，用于渲染其完整形状。 最终的 SVG $\mathcal{S}$ 通过遍历 $\mathcal{T}$，并根据 z 顺序（从背景到前景）约束嵌套图元而得到。

3.2. Top-Down VLM-Driven Decomposition

To construct $\mathcal{T}$, the pipeline executes a top-down recursive decomposition. The VLM acts as the core reasoning engine, converting visual complexity into a structured semantic ontology, which is then grounded into pixel space.

为了构建 $\mathcal{T}$，该流程执行自顶向下的递归分解。 VLM 作为核心推理引擎，将复杂的视觉内容转换为结构化语义本体，再将其定位到像素空间。

Hierarchical Semantic Decomposer Starting from the root image $\mathcal{I}$, we traverse the tree. Given an active node $v_i$, the VLM receives the root image and a textual prompt. Relying on its multimodal priors, it evaluates the semantic decomposability of $v_i$. If deemed decomposable, it generates a structured parsing plan: a sequence of child semantic labels $l_j$ and bounding boxes $b_j$ ordered back-to-front. If the node is evaluated as semantically atomic, it halts the descent and designates $v_i$ as a terminal leaf node $v_k\in \mathcal{L}$.

层级语义分解器 我们从根图像 $\mathcal{I}$ 出发遍历树。 给定当前处理的节点 $v_i$，VLM 接收根图像和文本提示作为输入。 它依据自身的多模态先验，评估 $v_i$ 在语义上是否可分解。 如果判定为可分解，它就生成结构化解析计划：按从后到前的顺序排列的一系列子节点语义标签 $l_j$ 和边界框 $b_j$。 如果该节点被判定为语义上不可再分，则停止向下展开，并将 $v_i$ 指定为终端叶节点 $v_k\in \mathcal{L}$。

Semantic Grounding To physically execute the VLM's decomposition plan, the root image and predicted constraints $(l_j,b_j)$ are passed to the Segment Anything Model (SAM). Here, SAM grounds the VLM's generated concepts, extracting pixel masks corresponding to the semantic labels.

语义定位 为了实际执行 VLM 的分解计划，我们将根图像和预测约束 $(l_j,b_j)$ 传递给 Segment Anything Model（SAM）。 在这里，SAM 将 VLM 生成的概念定位到图像中，提取与语义标签对应的像素掩码。

VLM Residual Judge Semantic grounding rarely achieves perfect pixel mutual exclusivity. We compile any unassigned pixels into a residual image and invoke the VLM as a Residual Judge. The VLM evaluates the residuals to determine if they constitute a missed semantic concept. If a valid semantic identity is identified, the residual is labeled and added to the tree as a new node; otherwise, it is discarded as non-semantic noise.

VLM 残差判别器 语义定位很少能够实现完全互斥的像素划分。 我们将所有未分配的像素汇集为残差图像，并调用 VLM 充当 残差判别器。 VLM 评估这些残差，以判断它们是否构成被遗漏的语义概念。 如果识别出了有效的语义身份，就为该残差赋予标签并将其作为新节点加入树中；否则，将其作为非语义噪声丢弃。

Recursive Tree Expansion This decomposition sequence is applied iteratively. The system dynamically expands $\mathcal{V}$ and $\mathcal{E}$ via Breadth-First Search (BFS), retaining the global image as a visual anchor while updating textual context at each depth level, continuing until all branches terminate in leaf nodes $\mathcal{L}$.

递归树扩展 上述分解流程会迭代执行。 系统通过广度优先搜索（BFS）动态扩展 $\mathcal{V}$ 和 $\mathcal{E}$，保留全局图像作为视觉参照，同时在每个深度层级更新文本上下文，直到所有分支都终止于叶节点 $\mathcal{L}$。

3.3. Prompt-Driven Semantic Completion

Extracting foreground components during top-down decomposition inherently leaves artificial “holes” in the background elements. To restore geometric integrity—which is absolutely vital to ensure the final SVG components can be independently edited by the user without revealing missing background data—the terminal leaf nodes $v_k\in \mathcal{L}$ undergo targeted occlusion recovery.

在自顶向下的分解过程中提取前景组件，必然会在背景元素中留下人为的“空洞”。 为了恢复几何完整性，终端叶节点 $v_k\in \mathcal{L}$ 会接受有针对性的遮挡恢复；这一点至关重要，因为它能确保用户独立编辑最终 SVG 组件时，不会暴露缺失的背景数据。

VLM Inpainting Judge Executing generative inpainting on every leaf is computationally inefficient. The VLM acts as an Occlusion Assessor, taking the isolated leaf image, root image, and text label $l_k$ as inputs. By relying on explicitly prompted geometric priors (e.g., instructions defining the silhouette of an occluder), the VLM classifies whether true occlusion exists.

VLM 修复判别器 对每个叶节点都执行生成式修复，在计算上并不高效。 VLM 充当 遮挡评估器，接收分离出的叶节点图像、根图像和文本标签 $l_k$ 作为输入。 它依据提示中明确给出的几何先验（例如定义遮挡物轮廓的指令），判断是否存在真实遮挡。

VLM-Guided Amodal Spatial Estimation For occluded leaves, the VLM functions as a Spatial Estimator. Given the leaf image and text label, it predicts a high-resolution 2D polygon (30–50 vertices) outlining the inferred amodal (complete) shape of the component.

VLM 引导的全形空间估计 对于被遮挡的叶节点，VLM 充当 空间估计器。 给定叶节点图像和文本标签，它预测一个高分辨率二维多边形（30–50 个顶点），勾勒出推断得到的组件全形（完整）形状。

Generative Inpainting The target inpainting mask is derived mathematically by subtracting the visible SAM mask from the VLM-generated amodal polygon. This exact difference mask, alongside the original leaf image and the textual label prompt, is passed to an inpainting model to recover the obscured geometry conditioned on the semantic concept $l_k$.

生成式修复 目标修复掩码通过从 VLM 生成的全形多边形中减去 SAM 得到的可见区域掩码，在数学上计算得到。 这一精确的差集掩码与原始叶节点图像、文本标签提示一起传入修复模型，在语义概念 $l_k$ 的条件下恢复被遮蔽的几何形状。

3.4. Semantic Markup Assembly

Once the visual components are fully isolated and recovered, an external vectorizer translates the pixel representations of the leaf nodes $\mathcal{L}$ into mathematical primitives. Finally, the tree $\mathcal{T}$ is compiled. Because SVG is natively an XML-based markup language, we construct the final output dynamically in a bottom-up pass. We nest SVG group tags (<g>) using the VLM-generated semantic labels as identifiers (id="label"). This assembly yields a final compositional SVG where the underlying text file perfectly mirrors the semantic hierarchy extracted by the VLM.

视觉组件被完全分离并恢复后，外部矢量化器将叶节点 $\mathcal{L}$ 的像素表示转换为数学图元。 最后，对树 $\mathcal{T}$ 进行编译。 由于 SVG 本身是一种基于 XML 的标记语言，我们通过一次自底向上的遍历动态构建最终输出。 我们嵌套 SVG 分组标签（<g>），并使用 VLM 生成的语义标签作为标识符（id="label"）。 这种组装方式产生了最终的组合式 SVG，其底层文本文件完整对应于 VLM 提取的语义层级。

3.5. Extension to Text-to-SVG

Although our framework is designed for the Image-to-SVG task, it naturally extends to Text-to-SVG by cascading a Text-to-Image (T2I) model with our pipeline. Given a text prompt $\mathcal{P}$, we synthesize a raster image ${\mathcal{I}}_{gen}$ via a T2I diffusion model and feed it directly into our pipeline as the root image $\mathcal{I}$. The semantic decomposition and grouping then proceed identically to the Image-to-SVG case.

尽管我们的框架是为图像到 SVG 任务设计的，但将文本到图像（T2I）模型与该流程串联，就能自然地扩展到文本到 SVG 任务。 给定文本提示 $\mathcal{P}$，我们利用 T2I 扩散模型合成光栅图像 ${\mathcal{I}}_{gen}$，并将其直接作为根图像 $\mathcal{I}$ 输入我们的流程。 随后，语义分解与分组过程便与图像到 SVG 的情形完全相同。

4. The Semantic SVG Benchmark

To establish a rigorous standard for semantic compositionality in SVG generation, we introduce the Semantic SVG Benchmark. Because existing datasets and metrics evaluate only whole-image visual fidelity, they are incapable of measuring whether an SVG is structured for intuitive downstream manipulation. Our benchmark explicitly tests for semantic-visual alignment, geometric integrity, and structural decomposability.

为了为 SVG 生成中的语义组合性建立严格的标准，我们提出了 Semantic SVG Benchmark。 由于现有数据集和指标只评估整幅图像的视觉保真度，它们无法衡量 SVG 是否具有支持直观下游操控的结构。 我们的基准明确检验语义与视觉的对齐、几何完整性以及结构可分解性。

Source Data and Subsampling To rigorously construct our benchmark, we collect candidate SVGs as broadly as possible from two streams. The first consists of evaluation benchmarks used in prior work, which span icons, complex illustrations, and emojis. The second consists of publicly available web sources, whose full list and licenses are provided in Appendix A.2.

数据来源与子采样 为了严谨地构建基准，我们从两类来源尽可能广泛地收集候选 SVG。 第一类是既有工作使用的评测基准，涵盖图标、复杂插画和表情符号。 第二类是公开可用的网络来源，其完整列表及许可证见 附录 A.2。

Rigorous Filtering Criteria To guarantee the quality of the ground-truth annotations, we apply a strict manual filtering process. We retain only images with a clearly decomposable semantic structure, focusing on samples with substantial occlusion and overlapping elements so that trivial decomposition is impossible. We also exclude automatically generated samples produced by color-cluster-based optimization, whose paths span multiple semantic regions and thus make a faithful ground-truth grouping ill-defined. More details are provided in Appendix A.3.

严格的筛选标准 为保证真实标注的质量，我们实施了严格的人工筛选流程。 我们只保留具有明确可分解语义结构的图像，重点选取存在显著遮挡和元素重叠的样本，使分解无法通过简单方式完成。 我们还排除了由基于颜色聚类的优化方法自动生成的样本，因为这些样本的路径横跨多个语义区域，导致无法明确界定忠实的真实分组。 更多细节见 附录 A.3。

Human Annotation Process Using a custom-built web interface, annotators group raw SVG paths into self-contained semantic entities (those that retain their identity when separated from the scene) and assign a textual label to each group. Starting from the root image, they recursively split each node until no meaningful decomposition remains, at which point the node is designated a leaf. More details are provided in Appendix A.3.

人工标注流程 标注者使用定制的 Web 界面，将原始 SVG 路径组成独立完整的语义实体（即从场景中分离后仍能保持自身身份的实体），并为每个分组指定文本标签。 从根图像开始，他们递归拆分每个节点，直到无法再进行有意义的分解，此时将该节点指定为叶节点。 更多细节见 附录 A.3。

Benchmark Statistics The finalized benchmark comprises 203 rigorously annotated test samples. The semantic trees exhibit a maximum depth of 2 and a minimum depth of 1 (representing direct children of the root), with a mean depth of 1.02.

基准统计 最终基准包含 203 个经过严格标注的测试样本。 这些语义树的最大深度为 2，最小深度为 1（表示根节点的直接子节点），平均深度为 1.02。

Inter-Annotator Agreement To verify the reliability of our annotations, each annotator additionally labeled the samples originally annotated by the other, yielding two independent semantic trees per sample. We measure the agreement between the two trees using our grouping-quality metrics, obtaining 0.0142 (MSE) and 0.9666 (DINO), which indicates near-complete agreement. The annotators agreed on the semantic structure in most cases, and disagreements were confined to whether a few individual paths, such as background or shadow elements, belong to a given group.

标注者间一致性 为了验证标注的可靠性，每位标注者还对另一位标注者最初负责的样本进行了标注，从而为每个样本得到两棵相互独立的语义树。 我们使用分组质量指标衡量两棵树之间的一致性，得到 0.0142（MSE）和 0.9666（DINO），表明它们几乎完全一致。 标注者在大多数情况下对语义结构达成了一致，分歧仅限于少数单独路径是否属于某个分组，例如背景或阴影元素。

4.1. Evaluating Semantic Grouping

Let $\mathcal{G}=\{g_1,\dots ,g_N\}$ denote the set of non-root GT nodes, and let $\mathcal{R}(\cdot )$ be the rasterization operator that renders a set of paths into an image. We score the agreement between two rendered regions with $s(\cdot ,\cdot )$, instantiated as either pixel-level MSE (lower is better) or DINO similarity (higher is better), computed over the union bounding box of the two renderings. Throughout, $\mathrm{best}$ denotes optimization of $s$ over the specified domain in the metric-appropriate direction: minimization for MSE and maximization for DINO similarity.

令 $\mathcal{G}=\{g_1,\dots ,g_N\}$ 表示所有非根真实（GT）节点的集合，令 $\mathcal{R}(\cdot )$ 表示将一组路径渲染为图像的光栅化算子。 我们使用 $s(\cdot ,\cdot )$ 为两个渲染区域之间的一致性评分，具体采用像素级 MSE（越低越好）或 DINO 相似度（越高越好），并在两幅渲染结果的联合边界框内进行计算。 全文中，$\mathrm{best}$ 表示在指定范围内沿适合该指标的方向优化 $s$：对于 MSE 取最小值，对于 DINO 相似度取最大值。

Given a generated SVG, let $\mathcal{Q}$ denote its complete set of drawable paths, with $|\mathcal{Q}|=K$. Each path in $\mathcal{Q}$ is treated as an atomic unit of grouping, and a candidate group $p\subseteq \mathcal{Q}$ is formed by jointly rendering a subset of these paths as a single component. We denote by $\mathcal{P}$ the set of candidate groups against which GT nodes are matched.

给定生成的 SVG，令 $\mathcal{Q}$ 表示其中全部可绘制路径的集合，且 $|\mathcal{Q}|=K$。 $\mathcal{Q}$ 中的每条路径都被视为分组的原子单位，候选分组 $p\subseteq \mathcal{Q}$ 则通过将这些路径的一个子集共同渲染为单个组件而形成。 我们用 $\mathcal{P}$ 表示与 GT 节点进行匹配的候选分组集合。

For methods that natively output semantic groups (predicted), $\mathcal{P}=\{p_1,\dots ,p_M\}$ is simply the set of predicted groups, excluding the root. Flat baselines, however, provide no such set. In this case (optimal), for each GT node $g_i$, we select the subset of generated paths that best reconstructs it:

对于原生输出语义分组的方法（predicted，预测分组），$\mathcal{P}=\{p_1,\dots ,p_M\}$ 就是排除根节点后的预测分组集合。 然而，扁平基线并不提供这样的集合。 在这种情况下（optimal，最优分组），对于每个 GT 节点 $g_i$，我们选择能够最佳重建该节点的生成路径子集：

$$\begin{array}{}\text{(1)}& \begin{array}{rl}{p}_i^{\star }& =\underset{p\subseteq \mathcal{Q}}{\arg \min }\\ & \mathrm{MSE}(\mathcal{R}(p),\mathcal{R}(g_i)).\end{array}\end{array}$$

We then use $p_i^{\star }$ as the representative group of that baseline for $g_i$, yielding $\mathcal{P}=\{p_1^{\star },\dots ,p_N^{\star }\}$. The exact solution defines the best GT-conditioned group that can be formed from the baseline's paths and therefore an upper bound on its grouping quality. In practice, we approximate this combinatorial search using the greedy algorithm detailed in Appendix B.2.

然后，我们将 $p_i^{\star }$ 用作该基线针对 $g_i$ 的代表分组，从而得到 $\mathcal{P}=\{p_1^{\star },\dots ,p_N^{\star }\}$。 精确解定义了能够从该基线的路径中形成的、以 GT 为条件的最佳分组，因此也是其分组质量的上界。 实际中，我们使用 附录 B.2 详述的贪心算法来近似这一组合搜索。

Semantic Recall We evaluate how well the GT semantic concepts are captured by the generated output, by scoring each GT node against the candidate group that best reconstructs it:

语义召回率 我们将每个 GT 节点与能够最佳重建它的候选分组进行评分，以评估生成结果对 GT 语义概念的覆盖程度：

$$\begin{array}{}\text{(2)}& \begin{array}{rl}\mathrm{Recall}& =\frac{1}{|\mathcal{G}|}\sum _{g_i\in \mathcal{G}}\underset{p_j\in \mathcal{P}}{\mathrm{best}}\\ & s(\mathcal{R}(p_j),\mathcal{R}(g_i)).\end{array}\end{array}$$

Semantic Precision Conversely, we evaluate the validity of the generated groups by scoring each of them against the GT node it best reconstructs, which penalizes arbitrary or semantically meaningless clusters:

语义精确率 反过来，我们将每个生成分组与它能够最佳重建的 GT 节点进行评分，以评估生成分组的有效性，从而惩罚任意或缺乏语义意义的聚类：

$$\begin{array}{}\text{(3)}& \begin{array}{rl}\mathrm{Precision}& =\frac{1}{|\mathcal{P}|}\sum _{p_j\in \mathcal{P}}\underset{g_i\in \mathcal{G}}{\mathrm{best}}\\ & s(\mathcal{R}(p_j),\mathcal{R}(g_i)).\end{array}\end{array}$$

Since $\mathcal{P}$ is constructed per GT node in the optimal setting, Precision is evaluated only in the predicted setting, where groups are produced independently of the GT.

由于在 optimal 设置下，$\mathcal{P}$ 是针对每个 GT 节点构造的，因此仅在分组独立于 GT 生成的 predicted 设置下评估精确率。

4.2. Evaluating Editability and Geometric Integrity

Merely predicting a grouped structure is insufficient; if those groups cannot be seamlessly manipulated by a user, the grouping itself is not useful. To measure this functional editability, we introduce the Post-Edit Rendering Error (PERE), which simulates the most basic editing operation, deleting an object, and examines what the edit leaves behind. A faithful deletion requires two properties at once: the removed group must contain exactly the paths of that object, and the geometry it had occluded must be recovered so that no hole is exposed.

仅仅预测出分组结构还不够；如果用户无法顺畅地操控这些分组，那么分组本身就没有用处。 为了衡量这种功能性可编辑性，我们提出编辑后渲染误差（PERE）：它模拟最基本的编辑操作——删除一个对象，并检验编辑后的剩余内容。 忠实的删除操作需要同时满足两个条件：被移除的分组必须恰好包含该对象的所有路径，并且它所遮挡的几何形状必须得到恢复，从而不留下暴露的空洞。

Let $\mathcal{E}(\mathcal{S},g)$ denote the image obtained by deleting group $g$ from an SVG $\mathcal{S}$ and rendering the result. Given a GT sample $\mathcal{S}$ with generated counterpart $\hat{\mathcal{S}}$, we enumerate the pairs of depth-1 GT groups whose amodal masks intersect, and take the front object of each pair as the deletion target, yielding a set of GT occluders $\mathcal{O}$. For each $o\in \mathcal{O}$, we locate its counterpart in the generated SVG by rendering every candidate group in isolation and taking the closest match,

令 $\mathcal{E}(\mathcal{S},g)$ 表示从 SVG $\mathcal{S}$ 中删除分组 $g$ 后渲染得到的图像。 给定 GT 样本 $\mathcal{S}$ 及其对应的生成结果 $\hat{\mathcal{S}}$，我们枚举全形掩码相交的深度为 1 的 GT 分组对，并将每一对中位于前方的对象作为删除目标，得到 GT 遮挡物集合 $\mathcal{O}$。 对于每个 $o\in \mathcal{O}$，我们单独渲染每个候选分组并选择最接近的匹配项，以在生成的 SVG 中定位其对应对象：

$$\begin{array}{}\text{(4)}& \begin{array}{rl}\hat{o}& =\underset{p\in \mathcal{P}}{\arg \min }\\ & \mathrm{MSE}(\mathcal{R}(p),\mathcal{R}(o)),\end{array}\end{array}$$

and then delete the two counterparts and compare the resulting images over the full frame:

然后删除这两个对应对象，并在整个图像范围内比较所得图像：

$$\begin{array}{}\text{(5)}& \begin{array}{rl}\mathrm{PERE}& =\frac{1}{|\mathcal{O}|}\sum _{o\in \mathcal{O}}\\ & \mathrm{MSE}(\mathcal{E}(\hat{\mathcal{S}},\hat{o}),\mathcal{E}(\mathcal{S},o)),\end{array}\end{array}$$

which we then average over the benchmark. Here $\mathcal{P}$ is the set of predicted groups for our method, and is defined as in Equation (1) for flat baselines. A high PERE indicates a practical failure in compositionality: the model either grouped unrelated paths together, causing unintended elements to disappear, or failed to recover the occluded geometry, exposing a hole where the object had been. Further details are provided in Appendix B.2.

随后在整个基准上对这一结果取平均。 这里，对于我们的方法，$\mathcal{P}$ 是预测分组的集合；对于扁平基线，其定义与 公式（1） 相同。 较高的 PERE 表明组合性在实际使用中失效：模型要么将不相关的路径分到一起，导致本不应删除的元素消失，要么未能恢复被遮挡的几何形状，使对象原来的位置暴露出空洞。 更多细节见 附录 B.2。

5. Experiments

5.1. Experimental Settings

Baselines For the Image-to-SVG task, we compare our method against dedicated VLM-based SVG generation models (OmniSVG-8B, InternSVG-8B, StarVector, LayerPeeler), a representative optimization-based baseline (VTracer), and general-purpose VLMs, namely open-source Qwen3.6-35B-A3B and proprietary Gemini-3-flash, both prompted with a dedicated instruction for semantic <g> grouping. For the Text-to-SVG task, we extend our pipeline with FLUX.1-dev and SD3.5-medium as the text-to-image models, fine-tuned with LoRA on our target domain, and compare against OmniSVG-8B, InternSVG-8B, and the same two general-purpose VLMs under the identical grouping prompts. Training details for the text-to-image models are provided in Appendix B.1.

基线 对于图像到 SVG 任务，我们将所提方法与专门的 VLM 驱动 SVG 生成模型（OmniSVG-8B、InternSVG-8B、StarVector、LayerPeeler）、具有代表性的优化基线 VTracer，以及通用 VLM 进行比较；后者包括开源的 Qwen3.6-35B-A3B 和闭源的 Gemini-3-flash，两者均使用了专门针对语义 <g> 分组的指令。 对于文本到 SVG 任务，我们使用通过 LoRA 在目标领域上微调的 FLUX.1-dev 和 SD3.5-medium 作为文本到图像模型来扩展流程，并与 OmniSVG-8B、InternSVG-8B 以及上述两种使用相同分组提示的通用 VLM 进行比较。 文本到图像模型的训练细节见 附录 B.1。

Evaluation We evaluate Image-to-SVG models on the Semantic SVG Benchmark using Semantic Recall and Precision (MSE and DINO), Post-Edit Rendering Error (PERE) and whole-image fidelity (MSE and DINO). In the optimal setting, flat baselines are evaluated using GT-conditioned post-hoc path selection, approximated by the greedy search detailed in Appendix B.2. For Text-to-SVG, we evaluate generation quality and text alignment on MMSVG-Bench using FID, CLIP, HPSv2, and Aesthetic scores.

评测 我们在 Semantic SVG Benchmark 上评估图像到 SVG 模型，使用的指标包括语义召回率与精确率（MSE 和 DINO）、编辑后渲染误差（PERE）以及整幅图像的保真度（MSE 和 DINO）。 在 optimal 设置中，我们通过以 GT 为条件的事后路径选择来评估扁平基线，并使用 附录 B.2 中详述的贪心搜索进行近似。 对于文本到 SVG 任务，我们在 MMSVG-Bench 上使用 FID、CLIP、HPSv2 和 Aesthetic 分数评估生成质量与文本对齐程度。

Implementation Details We employ Gemini-3-flash as the default VLM agent. For external vision tools, we use SAM 3 for semantic grounding, Flux-Fill for occlusion recovery, and VTracer for vectorization. All experiments are conducted on two NVIDIA Blackwell B200 GPUs. More implementation details are in Appendix B.

实现细节 我们使用 Gemini-3-flash 作为默认的 VLM 智能体。 在外部视觉工具方面，我们使用 SAM 3 进行语义定位、Flux-Fill 进行遮挡恢复、VTracer 进行矢量化。 所有实验均在两张 NVIDIA Blackwell B200 GPU 上进行。 更多实现细节见 附录 B。

5.2. Results

表1：使用分组指标评估图像到 SVG 任务的定量结果。optimal 表示通过事后 GT 匹配得到的理论上界（对扁平基线仅得到召回率），predicted 则评估原生生成的结构元数据。我们完全依靠预测得到的结构，在分组质量和功能性可编辑性方面超过了基线的最优上界，同时保持了极具竞争力的整图视觉保真度。最佳和次佳结果分别用粗体与下划线标出。

Methods	Grouping	Grouping quality (MSE)↓			Grouping quality (DINO)↑			Editability	Whole Image
		Recall	Precision	F1	Recall	Precision	F1	PERE↓	MSE↓	DINO↑
VTracer	optimal	.0227	N/A	N/A	.9392	N/A	N/A	.0400	.0036	.9922
OmniSVG	optimal	.0679	N/A	N/A	.8276	N/A	N/A	.1675	.1629	.7853
InternSVG	optimal	.0531	N/A	N/A	.8581	N/A	N/A	.0834	.0847	.8390
StarVector	optimal	.0790	N/A	N/A	.7730	N/A	N/A	.0969	.1156	.7480
LayerPeeler	optimal	.0295	N/A	N/A	.9040	N/A	N/A	.0398	.0339	.9148
Gemini-3-flash	optimal	.0389	N/A	N/A	.8983	N/A	N/A	.0754	.0670	.9139
	predicted	.0724	.0904	.0787	.8937	.8274	.8589	.0921	.0670	.9139
Qwen3.6-35B-A3B	optimal	.0699	N/A	N/A	.8471	N/A	N/A	.0991	.1070	.8409
	predicted	.0801	.0710	.0729	.7677	.7413	.7538	.0970	.1070	.8409
Ours	optimal	.0091	N/A	N/A	.9571	N/A	N/A	.0198	.0049	.9845
	predicted	.0174	.0195	.0155	.9487	.9426	.9453	.0210	.0049	.9845

图3：图像到 SVG 任务的定性结果。对于每个场景，我们展示整幅图像、分离的语义对象和剩余背景。GT 列中的整幅图像同时也是所有方法的输入图像。Ours（pred）表示我们的模型预测出的分组（图元集合）。对于其他生成结果，我们展示的是通过最优事后分组策略选取路径后渲染得到的语义对象。我们的方法能够有效分离独立实体，并通过高质量的全形修复恢复被遮挡的背景几何。

Grouping Quality Table 1 presents the quantitative results on the Image-to-SVG task, evaluated on our Semantic SVG Benchmark. Since flat generation baselines do not output semantic groups, we grant them their optimal path combination as a group so that Recall can at least be measured. While the two general-purpose VLMs do produce grouped outputs when explicitly prompted, their predicted groups fall well behind their own optimal bounds, indicating that they do not group accurately. Ours (optimal) achieves the best Recall in both MSE (0.0091) and DINO (0.9571), and Ours (predicted) relies solely on native predictions yet still surpasses the optimal configurations of all baselines across every grouping metric. This does not come at the cost of reconstruction quality: our whole-image fidelity is comparable to VTracer and far ahead of the generative baselines, while VTracer, despite its strong fidelity and the favorable optimal grouping granted to it, still falls short of our predicted groups in Recall (0.0227 vs. 0.0174 in MSE, 0.9392 vs. 0.9487 in DINO).

分组质量 表1 展示了图像到 SVG 任务在 Semantic SVG Benchmark 上的定量评测结果。 由于扁平生成基线不输出语义分组，我们将其最优路径组合作为分组，以使召回率至少能够得到衡量。 尽管两种通用 VLM 在得到明确提示时确实能输出分组结果，但其预测分组远逊于自身的最优上界，表明它们的分组并不准确。 Ours（optimal）在 MSE（0.0091）和 DINO（0.9571）的召回率上均取得最佳结果，而 Ours（predicted）仅依赖原生预测，却仍在所有分组指标上超过了全部基线的最优配置。 这并未以重建质量为代价：我们的整图保真度与 VTracer 相当，并远高于生成式基线；而 VTracer 尽管保真度很高，并且获得了有利的最优分组配置，其召回率仍不及我们的预测分组（MSE 为 0.0227 对 0.0174，DINO 为 0.9392 对 0.9487）。

Semantic Editability Beyond static structure, we assess whether the generated groups are practically usable through the Post-Edit Rendering Error (PERE). Our predicted groups reduce PERE by 47.5% relative to VTracer (0.0400 vs. 0.0210), remaining ahead of every baseline even under their optimal configurations. Flat-generation methods fail this dynamic evaluation because they merge distinct objects sharing a color into a single entangled path, or leave visible holes where an occluding component is removed. Our pipeline instead divides regions per object and recovers the occluded background through amodal inpainting, enabling clean downstream edits.

语义可编辑性 除静态结构外，我们还通过编辑后渲染误差（PERE）评估生成分组是否能在实际中使用。 相较于 VTracer，我们的预测分组将 PERE 降低了 47.5%（0.0400 对 0.0210），并且即使所有基线均采用各自的最优配置，我们仍然领先。 扁平生成方法无法通过这一动态评测，因为它们会把颜色相同的不同对象合并到一条相互纠缠的路径中，或者在移除遮挡组件后留下可见空洞。 我们的流程则按对象划分区域，并通过全形修复恢复被遮挡的背景，从而支持干净、准确的下游编辑。

Qualitative Results Figure 3 compares the whole image, isolated semantic object, and residual background (rows) across methods (columns). Removing the foreground tram exposes holes in the background geometry produced by VTracer, while most other baselines fail to reconstruct the target image coherently. VTracer also merges similarly colored regions from different objects into a single path. In contrast, our pipeline isolates components through semantic decomposition and recovers occluded geometry through amodal inpainting. Consequently, Ours (pred) closely matches the GT in both object isolation and background completeness, showing that native predictions alone support clean edits.

定性结果 图3 比较了各方法（列）在整幅图像、分离的语义对象以及剩余背景（行）上的结果。 移除前景中的有轨电车后，VTracer 生成的背景几何会暴露空洞，而大多数其他基线甚至无法连贯地重建目标图像。 VTracer 还会将不同对象中颜色相近的区域合并成一条路径。 相比之下，我们的流程通过语义分解分离组件，并通过全形修复恢复被遮挡的几何形状。 因此，Ours（pred）在对象分离和背景完整性方面均与 GT 高度一致，表明仅依靠原生预测就足以支持干净、准确的编辑。

表2：人工可编辑性研究。两名参与者在我们的方法和 VTracer 的输出上进行移动和删除编辑（各 15 个样本），每项编辑都由基于 GT SVG 构造的编辑前后目标指定。Avg. time 表示完成编辑的平均用时，Completed 表示未放弃且成功完成的编辑数。MSE 衡量编辑结果与真实数据目标图像之间的像素级误差，仅对已完成的编辑计算。最佳结果用粗体标出。

Edit	Model	Avg. time↓	Completed↑	MSE↓
Move	VTracer	37.4 s	11/15	0.057
	Ours	22.2 s	15/15	0.025
Remove	VTracer	15.5 s	13/15	0.023
	Ours	7.5 s	15/15	0.006

User Study on Semantic Editability While PERE quantifies editability by simulating structural edits, we further verify whether these structures actually help users perform edits. Table 2 reports a user study in which participants reproduce a target edit—moving or removing objects—on the outputs of our method and VTracer. The editing environment respects the group structure of each SVG, so that selecting a group manipulates all of its paths at once, while ungrouped paths must be selected individually; participants could give up when the edit seemed impossible. All edits on our SVGs were completed, roughly $1.9\times$ faster, since an object can be manipulated as a single unit. The gain is largest for removal, where amodal inpainting restores the occluded regions and yields a substantially lower MSE (0.006 vs. 0.023).

语义可编辑性的用户研究 PERE 通过模拟结构性编辑来量化可编辑性，而我们进一步验证这些结构是否确实能帮助用户完成编辑。 表2 报告了一项用户研究：参与者在我们的方法和 VTracer 的输出上复现目标编辑，即移动或删除对象。 编辑环境遵循每个 SVG 的分组结构，因此选择某个分组会同时操控其全部路径，而未分组的路径必须逐一选择；当编辑看起来无法完成时，参与者可以放弃。 在我们的 SVG 上，所有编辑均得以完成，完成速度约为 VTracer 的 $1.9\times$，因为一个对象可以作为一个整体来操控。 删除操作的收益最大：全形修复恢复了被遮挡区域，并显著降低了 MSE（0.006 对 0.023）。

表3：骨干 VLM 消融实验。我们评估了使用不同 VLM（Qwen3.6-35B-A3B、GPT-5、Claude-Sonnet-4.6 和 Gemini-3-flash）作为框架内结构推理智能体的影响。我们的方法直接在显式预测的分组结构上测量分组质量，而基线 VTracer 使用其最优路径组合进行评估。

Method	Backbone VLM	Grouping quality (MSE)↓			Grouping quality (DINO)↑			Editability	Whole Image
		Recall	Precision	F1	Recall	Precision	F1	PERE↓	MSE↓	DINO↑
VTracer	-	.0227	N/A	N/A	.9392	N/A	N/A	.0400	.0036	.9922
Ours	Qwen3.6-35B-A3B	.0666	.0535	.0555	.8541	.8641	.8576	.0787	.0123	.9588
	GPT-5	.0629	.0427	.0465	.8665	.8945	.8789	.0733	.0063	.9771
	Claude-Sonnet-4.6	.0457	.0417	.0402	.8869	.8727	.8788	.0491	.0076	.9717
	Gemini-3-flash	.0174	.0195	.0155	.9487	.9426	.9453	.0210	.0049	.9845

Impact of the Backbone VLM We evaluate the influence of the backbone VLM by comparing our default model, Gemini-3-flash, with Qwen3.6-35B-A3B, GPT-5, and Claude-Sonnet-4.6 in Table 3. Among the backbones, Gemini-3-flash performs best across all metrics and Claude follows, matching or exceeding most baselines in Table 1 and remaining comparable to the strongest baseline VTracer in PERE, even though its groups are natively predicted while VTracer is granted its optimal post-hoc grouping. GPT-5, in contrast, often under-segments and at times leaves the scene undecomposed altogether, so the pipeline vectorizes the input as a whole without amodal recovery, which explains its high whole-image fidelity despite poor grouping. Qwen's lower performance mainly stems from its limited box localization during decomposition, and this spatial inaccuracy cascades directly into functional editability. Such differences arise because the backbone must localize sub-parts before our pipeline converts them into editable groups, so structural quality naturally follows its spatial reasoning capability. This confirms that while our top-down strategy inherently improves structural integrity, a spatially precise agent such as Gemini-3-flash is essential for maximizing semantic atomicity.

骨干 VLM 的影响 我们在 表3 中将默认模型 Gemini-3-flash 与 Qwen3.6-35B-A3B、GPT-5 和 Claude-Sonnet-4.6 进行比较，以评估骨干 VLM 的影响。 在这些骨干模型中，Gemini-3-flash 在所有指标上表现最佳，Claude 紧随其后；Claude 达到或超过了 表1 中的大多数基线，并且在 PERE 上与最强基线 VTracer 相当，尽管前者使用原生预测分组，而后者获得了最优的事后分组。 相比之下，GPT-5 经常出现分割不足，有时甚至完全不分解场景，使得流程直接将整幅输入图像矢量化而不进行全形恢复，这也解释了为何它分组质量较差，却仍具有较高的整图保真度。 Qwen 的较低表现主要源于其在分解阶段的边界框定位能力有限，而这种空间定位误差会直接传递到功能性可编辑性上。 之所以出现这些差异，是因为骨干模型必须先定位子部件，我们的流程才能将其转换为可编辑分组，因此结构质量自然受其空间推理能力影响。 这证实了：虽然自顶向下的策略本身有助于改善结构完整性，但要尽可能实现语义原子性，仍然需要 Gemini-3-flash 这样空间定位精确的智能体。

表4：修复模块消融实验。我们评估了在语义分组过程中使用全形修复的影响。我们的方法直接在显式预测的分组结构上测量分组质量。

Inpainting	Grouping quality (MSE)↓			Grouping quality (DINO)↑			Editability	Whole Image
	Recall	Precision	F1	Recall	Precision	F1	PERE↓	MSE↓	DINO↑
✗	.0269	.0258	.0263	.9337	.9282	.9310	.0307	.0063	.9844
✓	.0174	.0195	.0155	.9487	.9426	.9453	.0210	.0049	.9845

Impact of Generative Occlusion Recovery We evaluate the necessity of our prompt-driven amodal inpainting module by comparing our pipeline with and without it in Table 4. When the inpainting module is disabled (✗), the pipeline still predicts semantic groups but fundamentally acts as a standard mask-and-crop system, strictly extracting only visible pixels. As a result, both grouping quality and functional editability deteriorate. Enabling generative occlusion recovery (✓) improves every metric, most notably reducing Grouping F1 in MSE by 41.1% and PERE by 31.6%. This demonstrates that recovering the full, unoccluded geometry of a semantic component ensures the generated SVG paths accurately align with the true physical structure of the ground-truth objects, rather than just their fragmented, visible remnants.

生成式遮挡恢复的影响 我们在 表4 中比较启用和禁用提示驱动全形修复模块的流程，以评估该模块的必要性。 禁用修复模块（✗）时，流程仍会预测语义分组，但本质上成为标准的掩码裁剪系统，严格地只提取可见像素。 因此，分组质量和功能性可编辑性都会下降。 启用生成式遮挡恢复（✓）后，所有指标均有所改善，其中最显著的是基于 MSE 的分组 F1 降低了 41.1%，PERE 降低了 31.6%。 这表明，恢复语义组件完整、未被遮挡的几何形状，能够确保生成的 SVG 路径准确对齐真实对象的物理结构，而不是仅对齐其零散的可见残片。

表5：文本到 SVG 任务的定量评测。按照 MMSVG-Bench 的协议，在其 300 条提示上进行评估。Ours（SD3.5）和 Ours（FLUX）分别表示使用 SD3.5-medium 和 FLUX.1-dev 作为文本到图像模型的流程。

Methods	FID↓	CLIP↑	HPSv2↑	Aesthetic↑
OmniSVG	138.91	0.2435	0.2328	4.558
InternSVG	146.18	0.2530	0.2343	4.592
Qwen3.6-35B-A3B	184.72	0.2900	0.2611	5.075
Gemini-3-flash	166.03	0.3060	0.2673	5.016
Ours (SD3.5)	157.91	0.3017	0.2657	5.051
Ours (FLUX)	171.72	0.3076	0.2732	5.248

Evaluation on Text-to-SVG Generation Table 5 presents the Text-to-SVG quantitative evaluation. Both of our variants substantially outperform the baselines, with FLUX achieving the best CLIP, HPSv2, and Aesthetic scores. While OmniSVG reports the lowest FID, this is an evaluation artifact: the ground-truth images used to compute FID are derived directly from its own training data, biasing the metric toward that distribution. Visual comparisons in Figure 4 further support these results.

文本到 SVG 生成的评测 表5 展示了文本到 SVG 的定量评测结果。 我们的两种变体都显著优于基线，其中 FLUX 取得了最佳的 CLIP、HPSv2 和 Aesthetic 分数。 虽然 OmniSVG 的 FID 最低，但这是由评测方式造成的现象：用于计算 FID 的真实图像直接来自其自身的训练数据，使该指标偏向这一数据分布。 图4 中的视觉比较进一步支持了这些结果。

OmniSVG and InternSVG fail entirely, and although Qwen and Gemini produce reasonable layouts, individual objects are often rendered with low fidelity. Both of our variants instead synthesize the scene faithfully. Furthermore, separating a generated SVG into its predicted foreground and residual background shows that the occluded geometry behind the foreground remains structurally complete, thanks to our prompt-driven amodal inpainting. These results show that our framework extends naturally to Text-to-SVG, producing outputs that are faithful to the prompt while retaining the same semantic grouping and geometric completeness.

OmniSVG 和 InternSVG 完全失败，而 Qwen 和 Gemini 虽然能生成合理的布局，但单个对象的渲染保真度往往较低。 我们的两种变体则都能忠实地合成场景。 此外，将生成的 SVG 分离为预测前景和剩余背景后可以看到，得益于提示驱动的全形修复，前景后方被遮挡的几何形状仍然保持结构完整。 这些结果表明，我们的框架可以自然地扩展到文本到 SVG 任务，在保持相同语义分组和几何完整性的同时，生成忠实于提示的结果。

图4：文本到 SVG 任务的定性结果。图中分别展示提示“一只猫睡在一摞书上”（上）和“折叠地图上方的放大镜”（下）所生成的 SVG。对于每个场景，我们展示整幅图像、分离的语义对象和剩余背景。OmniSVG 和 InternSVG 不生成语义分组，因此其对象行和剩余背景行留空。

6. Conclusion

In this work, we address a fundamental limitation of existing SVG generation methods, which produce flat, semantically agnostic collections of paths that are difficult to manipulate. To resolve this, we reframe the task as semantic compositional SVG generation, where a multi-role VLM recursively decomposes a scene into geometrically complete, language-aligned components. To evaluate this new paradigm, we introduce the human-annotated Semantic SVG Benchmark together with the sub-component metrics Semantic Recall/Precision and Post-Edit Rendering Error. Our pipeline natively generates semantic structures that surpass the optimal post-hoc groupings of existing baselines in both grouping quality and functional editability, while preserving state-of-the-art visual fidelity. We believe this establishes a foundation for producing practically manipulable vector graphics.

本文针对现有 SVG 生成方法的一项根本性局限展开研究：它们生成扁平、缺乏语义信息且难以操控的路径集合。 为解决这一问题，我们将任务重新表述为语义组合式 SVG 生成，让承担多种角色的 VLM 将场景递归分解为几何完整、与语言对齐的组件。 为了评估这一新范式，我们引入了人工标注的 Semantic SVG Benchmark，以及语义召回率／精确率和编辑后渲染误差等子组件指标。 我们的流程原生生成的语义结构，在分组质量和功能性可编辑性方面都超过了现有基线的最优事后分组，同时保持了最先进的视觉保真度。 我们认为，这为生成在实际中可操控的矢量图形奠定了基础。

Limitations

The main limitation of our framework is computational cost. On our 203-image curated benchmark, each image required on average 16.3 Gemini-3-flash calls (3.5 decompose $+$ 3.9 occlusion judge $+$ 7.1 polygon $+$ 1.8 pick), costing $0.21 per image (median $0.20; total $42.06). End-to-end latency averaged 255 s per image (median 254 s), of which 94.9% was spent on sequential FLUX.1-Fill-dev inference ($\sim 15$ GB bf16 GPU, 30 denoising steps at $1024\times 1024$, four polygon variants per occluded part) and only 5.0% on hierarchical decomposition. Since our pipeline is inherently modular, this bottleneck can be largely mitigated by asynchronous processing across stages.

我们框架的主要局限在于计算成本。 在精选的 203 张图像基准上，每张图像平均需要调用 Gemini-3-flash 16.3 次（3.5 次分解 $+$ 3.9 次遮挡判别 $+$ 7.1 次多边形预测 $+$ 1.8 次选择），每张图像成本为 0.21 美元（中位数为 0.20 美元；总计 42.06 美元）。 每张图像的平均端到端延迟为 255 秒（中位数为 254 秒），其中 94.9% 用于顺序执行 FLUX.1-Fill-dev 推理（bf16 模式下 GPU 显存占用为 $\sim 15$ GB，在 $1024\times 1024$ 分辨率下执行 30 步去噪，每个被遮挡部件使用四种多边形变体），只有 5.0% 用于层级分解。 由于我们的流程本身采用模块化设计，通过各阶段之间的异步处理，可以在很大程度上缓解这一瓶颈。