Source-linked AI summary
InternSVG: Towards Unified SVG Tasks with Multimodal Large Language Models
Haomin Wang, Jinhui Yin, Qi Wei, Wenguang Zeng, Lixin Gu, Shenglong Ye, Zhangwei Gao, Yaohui Wang, Yanting Zhang, Yuanqi Li, Yanwen Guo, Wenhai Wang, Kai Chen, Yu Qiao, Hongjie Zhang
TL;DR
SVG modeling is hindered by fragmented task coverage, limited datasets, and difficulty handling complex SVGs. InternSVG combines SAgoge, SArena, and a unified MLLM for understanding, editing, and generation. It reports consistent gains across tasks and benchmarks, including improvements over a leading proprietary baseline on SArena-Icon.
Problem
SVG modeling faces fragmented task coverage, limited dataset scale and diversity, and difficulty generalizing to long sequences and complex SVG content.
Source: §1 · p. 2Method
The paper combines the SAgoge dataset and SArena benchmark with InternSVG, a unified MLLM using SVG-specific tokens and two-stage training for understanding, editing, and generation.
Source: §1 · pp. 2–3Results
InternSVG consistently outperforms open and proprietary counterparts; on SArena-Icon it exceeds Claude-Sonnet-4 by about 11% higher acc, 34% higher PSNR, 56% lower FID, and 22% higher SSIM.
Source: §1 · p. 3, §6 · p. 10Takeaways & Limitations
SAgoge and SArena provide broad task coverage and standardized measurement for systematic investigation of general SVG modeling.
Source: §1 · p. 2
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
SVG modeling faces fragmented task coverage, limited dataset scale and diversity, and weak transfer across tasks. InternSVG combines a dataset, benchmark, and unified MLLM, with reported gains across understanding, editing, and generation.
1 INTRODUCTION
- Existing resources and methods have fragmented task coverage, limited scale and domain diversity, and difficulty generalizing to long sequences and complex SVG content.The introduction also notes that many approaches overlook SVG understanding and editing.Source: §1 · p. 2
- SAgoge combines static graphics and animations in about 16 million samples, covering understanding, editing, and generation across varied domains and difficulty levels.SArena standardizes tasks and metrics across icons, illustrations, chemical structures, and animations.Source: §1 · p. 2
- On SArena-Icon, InternSVG exceeds Claude-Sonnet-4 by about 11% higher acc, 34% higher PSNR, 56% lower FID, and 22% higher SSIM across four task types.The comparisons cover understanding, editing, Text-to-SVG, and Image-to-SVG, respectively.Source: §1 · p. 3
- InternSVG is a unified MLLM for SVG understanding, editing, and generation, using SVG-specific tokens and a two-stage training strategy.The reported evaluations show it outperforming traditional approaches and general-purpose open-source and proprietary models.Source: §1 · p. 3
2 RELATED WORKS
Prior SVG datasets and methods are constrained by narrow task coverage, limited scale, and weak generalization to complex content. InternSVG is presented as a unified MLLM spanning understanding, editing, and generation.
2.1 SVG DATASETS AND BENCHMARKS
- Existing SVG datasets range from narrow single-task resources to small broad benchmarks, while many omit professional domains, rich editing instructions, or image-conditioned generation.SVGenius covers understanding, editing, and generation but has about 2,400 queries; SAgoge is described as a larger, more diverse resource with approximately 16M task samples.Source: §2.1 · p. 3
2.2 SVG MODELING METHODS
- Earlier specialized and optimization-based methods captured low-level structure but struggled with semantic understanding, complex compositions, and computational cost.Examples include hierarchical VAEs and differentiable-rasterization methods such as DeepSVG, DiffVG, and LIVE.Source: §2.2 · p. 3
- Diffusion-based systems improve visual fidelity but offer restricted controllability and task diversity, while LLM-based systems mainly address generation and struggle with fragmented tasks and complex SVG content.InternSVG is framed as a unified MLLM that understands SVG semantics and structure, edits by instruction, and generates from visual or textual prompts.Source: §2.2 · p. 4
3 DATASET AND BENCHMARK
SAgoge supplies a large, diverse SVG dataset with task definitions and a construction pipeline, while SArena provides a curated benchmark and standardized metrics. The evaluation covers understanding, editing, generation, and animation tasks.
3.1 SAGOGE DATASET
- SAgoge contains more than 16 million training samples spanning icons, illustrations, chemical structures, and animations.Its construction pipeline gathers Internet-sourced and synthetic SVGs.Source: §3.1 · p. 4, §3.1.2 · p. 5
3.1.1 TASK DEFINITION
- SAgoge defines SVG description and question answering, ten icon-editing subtasks, and generation tasks for icons, illustrations, chemical structures, and animations.Icon editing includes eight low-level operations and two high-level tasks: semantic color editing and style transfer.Source: §3.1.1 · p. 4
3.1.2 DATASET CREATION PIPELINE
- The construction pipeline combines collected and synthetic data, normalizes SVGs to a 128×128 viewBox, and simplifies code while preserving core geometry, semantics, and hierarchy.It uses rendered images or videos, processed SVG code, and prompts for MLLM synthesis; understanding, generation, and editing data use distinct annotation procedures.Source: Figure 2 · p. 4, §3.1.2 · p. 5
3.2 SARENA BENCHMARK
- SArena aligns its task definitions and difficulty spectrum with SAgoge and includes icons, illustrations, chemical structures, and animation.Automated preprocessing and manual screening remove low-quality, corrupted, and semantically ambiguous files.Source: §3.2 · p. 5
3.2.1 EVALUATION METRICS
- Understanding is evaluated with multiple-choice accuracy, while editing and image-conditioned SVG tasks use rendered-output similarity metrics including DINO, SSIM, LPIPS, and PSNR.Text-to-SVG uses FID and FID-CLIP for distributional and semantic fidelity, plus CLIP-T2I for text alignment; animation tasks use FVD and ViCLIP-based scores.Source: §3.2.1 · pp. 5–6
- Unrenderable SVG outputs are replaced with pure black images or videos during metric computation to penalize invalid or incomplete generations.This evaluation rule accounts for both valid-output quality and generation robustness.Source: §3.2.1 · p. 6
4 METHOD
InternSVG combines SVG-aware tokens and subword-based initialization with a curriculum that moves from simpler SVGs to more complex examples. The design targets efficient representation and stable training across SVG tasks.
4.1 MODEL ARCHITECTURE
- InternSVG follows a ViT–MLP–LLM architecture with InternViT-300M as its vision encoder and Qwen2.5-7B as its language model.The model adds SVG-specific tokens and a tailored embedding initialization strategy.Source: §4.1 · p. 6
- The SVG tokenizer adds 55 tag tokens, 42 attribute tokens, and fine-grained number tokens, reducing sequence length and computational burden.The number vocabulary includes integers from -128 to 128 and 100 decimal tokens from .0 to .99.Source: §4.1 · p. 6
- Each SVG-specific token is initialized by averaging the pretrained-tokenizer embeddings of its constituent subwords.This approach aims to preserve semantic coherence and stabilize training rather than starting from random embeddings.Source: §4.1 · p. 7
- Subword-based initialization preserves vocabulary priors, lowers initial loss, and accelerates convergence during SVG-token adaptation.The reported effects include more stable early training and improved overall efficiency.Source: §4.1 · p. 7
4.2 TWO-STAGE TRAINING STRATEGY
- The two-stage curriculum addresses the imbalance between plentiful, simple icons and scarcer, longer, more complex illustrations.Training begins with short, simple Icon and Chemistry SVGs, then expands to all datasets and tasks after preliminary convergence; Icon data is resampled for balance.Source: §4.2 · p. 7
5 EXPERIMENTS
Experiments compare InternSVG with MLLMs and SVG-generation methods on SArena and prior benchmarks, reporting results across domains and task types. Ablations examine joint-task training, staged training, and SVG-specific tokens with subword initialization.
5.2 QUANTITATIVE EVALUATIONS
- InternSVG leads on SArena-Icon Overall accuracy, Illustration Text-to-SVG FID/FID-C, and Chemistry tasks; on Animation it surpasses open-source MLLMs and nears proprietary systems.The reported comparisons cover SArena-Icon, Illustration, Chemistry, and Animation.Source: Table 3 · p. 8, Table 4 · p. 8, §5.2 · pp. 8–9
5.1 EXPERIMENTAL SETUP
- The evaluation compares open-source MLLMs, proprietary systems, traditional SVG generators, and LLM-based approaches; training uses two stages and 96 NVIDIA A800 GPUs.The reported setup uses a per-device batch size of 1 and learning rates of 2 × 10−4 and 1 × 10−4 for stages one and two, respectively.Source: §5.1 · pp. 7–8, §5.2 · p. 8
5.3 QUALITATIVE VISUALIZATION
- Figure 4 examples are described as diverse and visually appealing, with clear structures and semantically rich content.The paper points to further visualizations and model comparisons in Appendix B.9.Source: Figure 4 · p. 9, §5.3 · p. 9
5.2 QUANTITATIVE EVALUATIONS
- On prior benchmarks, InternSVG tops SGP-Bench, scores 0.826 on UniSVG Final Score, and improves several SVG-Stack metrics using shorter code than StarVector.On SVG-Stack, it has superior Dino, LPIPS, and MSE scores with comparable SSIM, using 1.2k tokens versus StarVector’s 3.7k–5.3k.Source: §5.2 · p. 9
5.4 ABLATION STUDIES
- Joint training across understanding, editing, and generation consistently improves performance and achieves the best results across the evaluated task metrics.The study uses 100K samples per task grouping and reports gains in most metrics for three-task training.Source: §5.4 · pp. 9–10, Table 5 · p. 9
- Two-stage training improves illustration and animation generation; on long-sequence illustration generation, FID-C falls from 25.67 to 5.14 and DINO rises from 0.830 to 0.924.The comparison uses a single-stage model trained on the merged data as its baseline.Source: §5.4 · p. 10, Table 6 · p. 10
- Model T+E, combining SVG-specific tokens with subword initialization, performs best across both benchmarks, while token-equipped models have higher illustration-generation success rates.The authors relate the illustration trend to sequence compression from SVG-specific tokens.Source: §5.4 · p. 10, Table 7 · p. 10
6 CONCLUSIONS
The paper presents InternSVG as a unified data-benchmark-model suite and reports consistent improvements across SVG domains and tasks.
6 CONCLUSIONS
- InternSVG integrates SAgoge, SArena, and a unified MLLM to address SVG tasks spanning icons, scientific diagrams, long-sequence illustrations, and animations.The authors report consistent improvements in understanding, editing, and generation across domains and tasks.Source: §6 · p. 10
REPRODUCIBILITY STATEMENT
The paper states that its results can be reproduced using the provided resources and directs readers to the sections and appendices containing implementation and dataset details.
REPRODUCIBILITY STATEMENT
- Architecture and training settings are in Sections 4–5; token implementation and dataset construction details are in the appendices and Section 3.1.The paper states that all presented results can be reproduced with the provided resources.Source: REPRODUCIBILITY STATEMENT · p. 11
ETHICS STATEMENT
The authors report that the work uses publicly available sources and that they do not anticipate adverse societal impact from their methods or datasets.
ETHICS STATEMENT
- The authors report excluding content they considered discriminatory or potentially in violation of copyright terms.They state that dataset elements came from open-source websites and publicly available sources, and that they do not anticipate adverse societal impact.Source: ETHICS STATEMENT · p. 11
A IMPLEMENTATION DETAILS OF INTERNSVG
InternSVG normalizes SVG coordinates and extends the vocabulary with specialized tokens initialized from existing subwords. The evaluation compares proprietary and open-source MLLMs, traditional SVG generators, and LLM-based SVG generation methods.
A IMPLEMENTATION DETAILS OF INTERNSVG
- SVGs are normalized to a canonical 128 × 128 coordinate system to reduce coordinate drift and computational burden.The authors state this also mitigates coordinate hallucination and enhances cross-task consistency.Source: A IMPLEMENTATION DETAILS OF INTERNSVG · p. 15
- The extended vocabulary includes 55 tag tokens, 42 attribute tokens, and compact integer and fractional tokens for numerical values.The numeric inventory comprises 247 integer tokens spanning −128 to 128, plus 100 two-decimal and 10 one-decimal fractional tokens; it is intended to shorten sequences and reduce computational requirements.Source: A IMPLEMENTATION DETAILS OF INTERNSVG · p. 15
- Added-token embeddings average the original tokenizer’s constituent subword embeddings, and training then proceeds end-to-end without parameter freezing.The authors state that this initialization stabilizes optimization, accelerates convergence, and improves grammatical decoding and numerical fidelity.Source: A IMPLEMENTATION DETAILS OF INTERNSVG · p. 15
B.1 EVALUATED MODELS
- The evaluation includes proprietary and open-source MLLMs, traditional SVG generation methods, and LLM-based SVG generation approaches.The listed comparison groups include GPT-4o and Gemini-2.5-Flash; Llama, Qwen, GLM, and InternVL models; VectorFusion and SVGDreamer; and LLM4SVG, StarVector, and OmniSVG.Source: §B.1 · p. 15
B.2 COMPLETE EVALUATION RESULTS ON SARENA
On SArena, InternSVG reports strong results across SVG understanding, editing, generation, and animation. Results on SGP-Bench, SVG-Stack, and UniSVG also show strong performance beyond SArena.
B.2 COMPLETE EVALUATION RESULTS ON SARENA
- InternSVG achieves the best results across SVG understanding metrics, surpassing Claude-Sonnet-4 by 8 points in Overall accuracy, and performs strongly on simple and hard editing tasks.It achieves perfect scores in several simple editing subtasks and leads in hard semantic color editing, but trails Claude-Sonnet-4 slightly on style transfer.Source: §B.2 · pp. 15–17, Table 9 · p. 16, Table 10 · p. 17
- On SArena-Icon and SArena-Illustration, InternSVG consistently outperforms prior SVG-generation approaches across major metrics and achieves the best Image-to-SVG performance across all metrics.Against large open-source models, it has lower FID and FID-C and higher CLIP-I2I scores, with a slight gap on CLIP-T2I.Source: §B.2 · p. 17, Table 12 · p. 18, Table 13 · p. 19
- On SArena-Chemistry, InternSVG achieves the lowest Text-to-SVG FID (9.974) and FID-C (0.877), highest CLIP-I2I (93.931), and best Image-to-SVG results on DINO (0.994), SSIM (0.873), and PSNR (17.722).The instructions use common or IUPAC compound names to guide structural-formula generation.Source: §B.2 · p. 18, Table 14 · p. 19
B.3 RESULTS ON SGP-BENCH
- On SGP-Bench, InternSVG reaches 72.3% overall accuracy, 6.2 percentage points above the second-best open-source MLLM and 1.3 points above Claude-Sonnet-4.It also leads on the Count and Shape subtasks.Source: §B.3 · pp. 19–20, Table 16 · p. 20
B.4 RESULTS ON SVG-STACK
- On SVG-Stack, InternSVG achieves the best performance on Dino, LPIPS, and MSE while using 1.2k output tokens versus StarVector’s 3.7k–5.3k, with comparable SSIM.Source: §B.4 · p. 20, Table 17 · p. 21
B.5 RESULTS ON UNISVG
- InternSVG achieves a UniSVG Final Score of 0.826, outperforming all baselines; it also leads on Image-to-SVG and Text-to-SVG results and Hard-set understanding accuracy.The Image-to-SVG ISVGEN score is 0.939, a 0.166 improvement over the second-best method; the Text-to-SVG TSVGEN score is 0.741, and Hard-set understanding accuracy is 0.625.Source: Table 18 · p. 21, §B.5 · p. 21
B.6 QUANTITATIVE EVALUATION OF SVG CODE QUALITY
The section evaluates SVG code structure, training strategy, cross-domain transfer, and qualitative outputs. The reported results cover concise and fast SVG generation, benefits of two-stage training, and visual comparisons across task domains.
B.6 QUANTITATIVE EVALUATION OF SVG CODE QUALITY
- InternSVG produces an average of 4.48 paths and generates outputs in 6.9 seconds, compared with 256.00 paths and 289 seconds for SVGDreamer and 64.00 paths and 71 seconds for VectorFusion.The authors also report that InternSVG supports the full set of SVG primitives, while optimization-based methods use heavy Bézier-curve approximations.Source: Table 19 · p. 21, §B.6 · pp. 21–22
B.7 TWO-STAGE VS. SINGLE-STAGE TRAINING
- Two-stage training is comparable on icon generation and superior on FID, FID-C, and SSIM, while consistently outperforming one-stage training on icon editing and chemistry, illustration, and animation generation.The authors report that the progressive strategy alleviates data imbalance and improves SVG modeling performance.Source: Table 20 · p. 22, §B.7 · p. 22
B.8 CROSS-DOMAIN GENERALIZATION ANALYSIS
- An InternSVG variant trained only on icon Image-to-SVG data improves over InternVL3-8B across all reported metrics on SArena-Chemistry Image-to-SVG.The zero-shot transfer results are DINO 0.865 to 0.930, SSIM 0.783 to 0.824, LPIPS 0.203 to 0.156, and PSNR 13.84 to 16.296.Source: §B.8 · p. 22, Table 21 · p. 22
B.9 ADDITIONAL QUALITATIVE VISUALIZATION
- For icon and illustration Text-to-SVG, InternSVG outputs are described as high-quality and semantically aligned, while competing methods show instruction-following, structural-stability, or validity problems.The visual comparisons also report better Image-to-SVG performance than general-purpose MLLMs, LLM-based SVG models, and LIVE, with fewer tokens and stronger detail preservation.Source: §B.9 · p. 23, Figure 5 · p. 23, Figure 7 · p. 25
- Qualitative comparisons report that InternSVG handles chemical-bond details better than StarVector and produces animations with higher fidelity and more consistent temporal dynamics.Source: §B.9 · p. 23, Figure 9 · p. 27, Figure 10 · p. 27
B.10 USER STUDY AND EVALUATION DETAILS
The appendix reports human evaluations of synthetic annotations and SVG quality, comparisons of SVG-generation models, and statistics for SAgoge and SArena.
B.10 USER STUDY AND EVALUATION DETAILS
- The human evaluation collected 42 valid responses and sampled outputs across icons, illustrations, chemical structures, and animations for six-model ratings on four dimensions.Participants rated semantic match or visual similarity, visual quality, human design alignment, and usability on a 1–5 scale.Source: §B.10 · pp. 33–34
- 95.04% overall semantic consistency was reported for synthetic instructions; participants preferred model-generated instructions in 45.43% of cases, versus 15.67% for human-authored instructions.36.90% of samples were considered equally good, and more than 82% were judged comparable to human annotations in quality.Source: §B.10 · p. 34, Table 22 · p. 34
- Illustrations were 96.19% defect-free (Human-Likeness Score 4.20; Code Quality Score 4.31), versus animations at 93.81%, 4.07, and 4.23, respectively.The study assessed visible defects, perceived likelihood of human authorship, and SVG code quality and readability on 1–5 scales.Source: §B.10 · p. 34, Table 23 · p. 34
- InternSVG significantly outperforms baseline models in instruction adherence and image reconstruction fidelity, consistent with quantitative metrics.For Image-to-SVG icons, user evaluation scores were InternSVG (4.64), Claude-Sonnet-4 (4.00), OmniSVG-3B (3.90), Gemini-2.5-Flash (3.73), GPT-4o (3.72), and Qwen2.5-VL-72B (2.42).Source: Figure 17 · p. 35, §B.10 · p. 36
C.1 STATISTICS OF SAGOGE
- SAgoge spans four domains: icons (2.8M SVGs; 11M samples), illustrations (600k; 1.6M), animations (61K; 122K), and chemistry (1.7M; 3.4M).Average sequence lengths are 846 tokens for icons, 8.7k for illustrations, and 1.7k for chemistry; animations include dynamic elements capturing temporal changes.Source: §C.1 · p. 37, Table 24 · p. 37
C.2 STATISTICS OF SARENA
- SArena covers Icon, Illustration, Animation, and Chemistry domains, with multiple-choice QA, editing, text-to-SVG, image-to-SVG, and video-to-SVG tasks.Avg. Tokens denotes the average token length of the training sample.Source: §C.2 · p. 37, Table 25 · p. 37
C.3.1 ILLUSTRATION DATA SYNTHESIS PIPELINE
The section describes dataset-construction pipelines for static illustrations and animations, using image generation and vectorization for illustrations and constrained SMIL generation for animations. It also outlines detailed appearance-and-composition templates and task-specific prompts for SVG understanding, editing, and generation.
C.3.1 ILLUSTRATION DATA SYNTHESIS PIPELINE
- GPT-4o prompts guide a FLUX-based model with a vector-style LoRA, then VTracer converts generated images into SVGs by detecting paths, curves, and fills.The pipeline targets prompt consistency and SVG-like features including simplified shapes, flat colors, and clean edges, enabling systematic production of diverse illustrations.Source: §C.3.1 · p. 38
C.3.2 ANIMATION DATA SYNTHESIS PIPELINE
- Claude-Sonnet-4 synthesizes SMIL-compliant SVG animations using fixed canvas sizes and mandatory animation primitives to support structural consistency, renderability, and dataset scale and diversity.The pipeline expands scarce open-source animation resources and produces animations with richer semantic content and greater geometric diversity.Source: §C.3.2 · p. 38
C.4.1 TEMPLATE FOR DATASET CONSTRUCTION
- Construction examples describe illustrations through depicted objects, colors, shapes, strokes, and composition.Examples include a woman gesturing toward a key and lock, with flat fills and an asymmetric layout, and a centrally aligned chocolate mousse cup with rounded shapes and no visible strokes.Source: §C.4.1 · p. 53
C.4.2 QUESTION TEMPLATE
- SVG description prompts vary wording while asking for semantic meaning and geometric features; multiple-choice QA prompts require only a letter-form answer.Description questions can ask about objects, shapes, layout, and symbolic cues, while QA instructions request a final answer such as “A.” without explanation.Source: §C.4.2 · pp. 53–54
- SVG editing prompts require minimal modifications while preserving code structure, IDs, and classes, and returning only the raw edited SVG.The templates discourage refactoring or reformatting and request the complete modified code without commentary or markdown.Source: §C.4.2 · p. 54
- Text-to-SVG prompts use varied dialogue formulations to elicit SVG code that fulfills a textual instruction.The templates frame the model as an SVG expert or designer and request an SVG matching the instruction.Source: §C.4.2 · p. 55
- Image-to-SVG and text-to-SANI prompts condition SVG generation on an image and instruction or on a textual instruction, respectively.The image-to-SVG templates provide both inputs; the animation templates request an SVG animation from text.Source: §C.4.2 · pp. 55–56
- Video-to-SANI prompts condition SVG animation generation on both a video and a textual instruction.The templates repeatedly specify the video input alongside the instruction to guide the animation output.Source: §C.4.2 · pp. 56–57