WildSVG: Towards Reliable SVG Generation Under Real-Word Conditions

Marco Terral; Haotian Zhang; Tianyang Zhang; Meng Lin; Xiaoqing Xie; Haoran Dai; Darsh Kaushik; Pai Peng

cs.CVFeb 24, 2026

WildSVG: Towards Reliable SVG Generation Under Real-Word Conditions

Authors: Marco Terral, Haotian Zhang, Tianyang Zhang, Meng Lin, Xiaoqing Xie, Haoran Dai, Darsh Kaushik, Pai Peng,

Organizations: QuiverAI · Columbia University · Illinois Institute of Technology · Mila - Quebec Artificial Intelligence Institute ·

Abstract

We introduce the task of SVG extraction, which consists in translating specific visual inputs from an image into scalable vector graphics. Existing multimodal models achieve strong results when generating SVGs from clean renderings or textual descriptions, but they fall short in real-world scenarios where natural images introduce noise, clutter, and domain shifts. A central challenge in this direction is the lack of suitable benchmarks. To address this need, we introduce the WildSVG Benchmark, formed by two complementary datasets: Natural WildSVG, built from real images containing company logos paired with their SVG annotations, and Synthetic WildSVG, which blends complex SVG renderings into real scenes to simulate difficult conditions. Together, these resources provide the first foundation for systematic benchmarking SVG extraction. We benchmark state-of-the-art multimodal models and find that current approaches perform well below what is needed for reliable SVG extraction in real scenarios. Nonetheless, iterative refinement methods point to a promising path forward, and model capabilities are steadily improving

Figures & tables

Figure 1: Performance on WildSVG . Extracting the logo from a real image and converting it to SVG using a two step pipeline. Comparison of VLMs on the WildSVG natural dataset.

Figure 2: Creation of the WildSVG Benchmark and evaluation pipeline . The dataset combines a synthetic split, built by rendering real SVGs into generated scenes, and a natural split, created by detecting logos in real images and pairing them with SVG annotations. We evaluate SVG extraction using one step and two step multimodal methods and assess outputs through pixel similarity, semantic similarity, code quality, and editability metrics.

Name	Classes	Images	Objects	Origin and Types
Belgalogos [ 19 ]	37	10,000	2,695	Photojournalist archives, general logos
FlickrLogos-27 [ 20 ]	27	1,080	4,671	General logos from Flickr search
FlickrLogos-32 [ 32 ]	32	2,240	5,644	General logos from Flickr search
SportLogo [ 22 ]	31	2,836	—	NHL and NBA logos from web search
Logos in the Wild [ 36 ]	871	11,054	32,850	Logos captured in diverse real scenes
QMUL OpenLogo [ 34 ]	352	27,083	—	Aggregation of 7 logo detection datasets

Table 1: Overview of publicly available logo detection datasets. Existing datasets cover a wide range of logo classes, domains and scales, yet all operate purely in the raster domain . To the best of our knowledge, no dataset provides paired SVG annotations , highlighting the gap that WildSVG aims to fill.

Dataset	Train	Val	Test	Primitives	Annotations
SVG Stack	2.1M	108k	5.7k	All	Captions
SVG Diagrams	—	—	472	All	Captions
SVG Fonts	1.8M	91.5k	4.8k	Vector paths	Font type
SVG Emoji	8.7k	667	668	All	Class
SVG Icons	80.4k	6.2k	2.4k	Vector paths	Class, caption

Table 2: Overview of public SVG generation datasets. Existing datasets provide large collections of SVGs across icons, fonts and diagrams, but none include natural image contexts or real world SVG extraction settings.

Dataset	Train / Val / Test	Primitives	Annotations
Natural	12759 / 1418 / 227	Path	Logo brand, focus prompt, description, bounding box
Synthetic	2104 / 190 / 99	All	Focus prompt, description, bounding box (test only)

Table 3: WildSVG includes Natural samples where real images contain visible logos that are matched to their SVG versions, and Synthetic samples created by using image generation models to place each input SVG into a generated scene. The table reports sample counts, primitives and annotations.

Figure 3: Examples of Synthetic WildSVG . Real SVGs extracted from SVG Stack are integrated into realistic scenarios using an image generation model.

Figure 4: Examples of Natural WildSVG . Real images with visible logos are associated with the SVG logos that are extracted from public logo databases.

Figure 5: Qualitative comparison of VLM outputs on the two step SVG extraction task using the Synthetic split . Given the cropped logo from the detector, each model produces an SVG reconstruction. Results illustrate varying levels of geometric fidelity, color consistency, and structural correctness across models.

One-step generation
Model	Natural						Synthetic
Model	DINO ↑	LPIPS ↓	L2 ↓	SSIM ↑	Len diff ↓	Prim div ↑	DINO ↑	LPIPS ↓	L2 ↓	SSIM ↑	Len diff ↓	Prim div ↑
Qwen2.5VL 8B SVG RL	0.69	0.39	0.15	0.63	8108	2.57	0.76	0.43	0.16	0.61	10375	3.12
Qwen2.5VL 72B Instruct	0.77	0.41	0.22	0.58	8804	2.15	0.77	0.42	0.21	0.58	1516	1.86
GLM 4.5V	0.79	0.39	0.18	0.61	8710	2.02	0.77	0.40	0.19	0.59	1371	1.95
VCode Claude Sonnet 4.5	0.78	0.46	0.22	0.53	8141	3.34	0.85	0.38	0.20	0.59	920	2.66

Table 4: Main results on WildSVG. We report visual similarity metrics (DINO, LPIPS, L2, SSIM) and code metrics (length difference and primitive diversity) for one-step and two-step SVG extraction on the Natural and Synthetic splits. Lower is better for LPIPS, L2, and length difference. Higher is better for DINO, SSIM, and primitive diversity.

Figure 6: Comparison of VLMs for one-step SVG extraction natural dataset

Figure 7: Comparison of VLMs for one-step SVG extraction synthetic dataset

Figure 8: Comparison of VLMs for two-step SVG extraction synthetic dataset

One-step generation
Model	Natural						Synthetic
Model	DINO ↑	LPIPS ↓	L2 ↓	SSIM ↑	Len diff ↓	Prim div ↑	DINO ↑	LPIPS ↓	L2 ↓	SSIM ↑	Len diff ↓	Prim div ↑
Qwen2.5VL-72B-Instruct	0.77	0.41	0.22	0.58	8803.83	2.15	0.77	0.42	0.21	0.58	1516.36	1.86
Qwen2.5VL-8B SVG RL	0.69	0.39	0.15	0.63	8108.38	2.57	0.76	0.43	0.16	0.61	10374.72	3.12
Gemini Flash 2	0.78	0.38	0.17	0.61	8233.99	1.69	0.83	0.32	0.19	0.63	1202.83	1.51
Gemini Flash 2.5	0.79	0.42	0.20	0.58	8509.60	2.11	0.78	0.43	0.21	0.57	192.90	1.82

Table 5: Complete unified results on WildSVG. We include the diverse families and versions benchmarked on WildSVG, including both the one-step and two-step SVG extraction on the Natural and Synthetic splits.

Figure 9: Pipeline for synthetic WildSVG generation

Figure 10: Pipeline for natural WildSVG generation

Figure 11: Prompt for image generation of Synthetic WildSVG Dataset

Figure 12: Prompt for scoring Natural WildSVG instances

Figure 13: Prompt for scoring Synthetic WildSVG instances

Figure 14: Bounding boxes position heatmap for WildSVG dataset test split

Figure 15: Distribution of bounding boxes coverage of the image for WildSVG dataset test split

Figure 16: Score distribution for WildSVG dataset test split

Figure 17: Comparison of VLMs for one-step SVG extraction natural dataset

Figure 18: Comparison of VLMs for one-step SVG extraction synthetic dataset

Figure 19: Comparison of VLMs for two-step SVG extraction natural dataset

Figure 20: Examples of real-life images and associated SVG for natural WildSVG

Figure 21: Examples of real-life images and associated SVG for synthetic WildSVG

Vector Graphics Generation

See all 26 papers

SVGEval: A Vision-Grounded Framework for Perceptual-Quality Benchmarking and Evaluation in Text-to-SVG Generation
Aug 3, 2026Yiming Wang, Ye Chen, Hanqi Chen +1Multimodal Model Evaluation Text-to-SVG Generation
School of Integrated Circuits (School of Information Science and Electronic Engineering), Shanghai Jiao Tong University, Shanghai, China · Zhiyuan College, Shanghai Jiao Tong University, Shanghai, China · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen 518107, China
VectorGym: A Multi-Task Benchmark for SVG Code Generation, Sketching and Editing
Feb 22, 2026Joan Rodriguez, Haotian Zhang, Abhay Puri +14VLM Evaluation Code Generation
Mila, Quebec AI Institute · ÉTS Montréal · QuiverAI +11
RefineSVG: Visual Feedback-Driven Reinforcement Learning for Image-to-SVG Generation
Jul 30, 2026Shaobo Liu, Feiqiao Mao, Shuaishuai Zhou +4Image Generation Large Vision-Language Models
Shenzhen University · Shenzhen, China · Peking University +1

Aug 3, 2026J/K move · Enter open · S save

SVGEval: A Vision-Grounded Framework for Perceptual-Quality Benchmarking and Evaluation in Text-to-SVG Generation

Yiming Wang, Ye Chen, Hanqi Chen, Bingbing Ni

School of Integrated Circuits (School of Information Science and Electronic Engineering), Shanghai Jiao Tong University, Shanghai, China · Zhiyuan College, Shanghai Jiao Tong University, Shanghai, China · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen 518107, China

Multimodal Model Evaluation Text-to-SVG Generation

Open paper AskPDF arXiv

Multimodal large models are increasingly used to generate scalable vector graphics (SVG), but reliable evaluation remains underexplored. Existing protocols are often code-centric or borrow raster-image metrics after rendering SVGs, which fail to reflect human perception and overlook SVG-specific qualities such as geometry and spatial composition. We introduce SVGEval, a vision-grounded multimodal benchmark for human-aligned SVG quality assessment. SVGEval explicitly incorporates visual renderings to evaluate whether models can judge the rendered outcome rather than only inspect SVG code, and provides high-quality annotations obtained via multi-round human labeling with expert refinement. Systematic evaluations across representative multimodal models reveal a clear gap: models perform relatively well on semantic alignment and aesthetics, yet struggle on geometry- and layout-related judgments. Building on SVGEval, we train an explainable SVG quality scorer that outputs multi-aspect scores with textual rationales. Ablations show that explicit visual grounding and reasoning supervision are crucial, especially for spatial and geometric assessment. SVGEval offers a reliable testbed and practical scorer for evaluating and improving SVG generation in the era of multimodal models.

Abstract

Vector Graphics Generation

Explore similar work

SVGEval: A Vision-Grounded Framework for Perceptual-Quality Benchmarking and Evaluation in Text-to-SVG Generation

VectorGym: A Multi-Task Benchmark for SVG Code Generation, Sketching and Editing

RefineSVG: Visual Feedback-Driven Reinforcement Learning for Image-to-SVG Generation