UniSVG: Multimodal SVG Benchmark

• UniSVG is an SVG-centric dataset defined by its unique focus on vector graphics as executable, editable code, emphasizing both geometric fidelity and syntactic validity.
• It supports unified tasks including image-to-SVG, text-to-SVG, and SVG understanding, with evaluations using metrics such as SSIM, LPIPS, and CLIP similarity.
• The benchmark is built on over 525K items and employs rigorous data cleaning and a balanced task taxonomy to drive advancements in multimodal large language model performance.

Searching arXiv for papers on UniSVG and related SVG generation/parsing benchmarks. UniSVG is an SVG-centric dataset and benchmark for unified SVG understanding and generation with Multimodal LLMs (MLLMs). It is designed to make image-to-SVG, text-to-SVG, and SVG understanding tractable within a single multimodal setting, using SVG code, raster renderings, and text supervision as aligned representations. The benchmark treats SVG as executable, editable code rather than as a raster target, so fidelity depends not only on semantic correctness but also on geometric precision, topology, and syntactic validity; later work has consequently used UniSVG both as a core benchmark for image-to-code evaluation and as a reference point for broader multimodal parsing and reinforcement-learning systems (Li et al., 11 Aug 2025, Zheng et al., 30 Jun 2026).

1. Conceptual scope and problem formulation

UniSVG addresses a distinctive regime of multimodal generation in which the output is a vector program. Unlike raster images, SVG encodes geometry, paths, transforms, and styling in XML-like code with floating-point parameters and retains resolution independence. This makes SVG attractive for computer vision, design, and program synthesis, but also makes the task unusually brittle: small numeric perturbations in coordinates, control points, or transforms can produce large visual deviations.

The benchmark formalizes both generation and understanding around this code-centric view. In the image-to-SVG setting used by later evaluators, the task can be stated as follows: given an input image $x$ and, optionally, a brief instruction $I$, a model must produce SVG code $y$ whose rendering $R(y)$ matches $x$ with high geometric, structural, and stylistic fidelity. This differs from ordinary image generation because the target is a syntactically valid program with explicit element structure such as <path>, <rect>, <circle>, gradients, fill and stroke styles, transforms, and coordinates (Zheng et al., 30 Jun 2026).

The difficulty is rooted in the geometry of vector graphics. UniSVG’s technical framing emphasizes Bézier curves, arcs, and affine transforms as first-class objects of prediction. For cubic Bézier curves with control points $P_0, P_1, P_2, P_3$, the curve is

$$B(t) = (1 − t)^3 P_0 + 3(1 − t)^2 t P_1 + 3(1 − t) t^2 P_2 + t^3 P_3,\quad t \in [0, 1].$$

Quadratic Bézier curves and SVG arcs are likewise parameterized continuously, and affine transforms act directly on coordinates. This parameter sensitivity is a central reason UniSVG is harder than raster-conditioned captioning or semantic image retrieval: semantic plausibility is insufficient if topology or geometry is wrong (Li et al., 11 Aug 2025).

2. Dataset construction, scale, and task taxonomy

UniSVG comprises 525,741 training items and a 2,850-item held-out benchmark. It is built from 360k cleaned SVGs and augmented with images and textual descriptions to support three task families: Image-to-SVG Generation (ISVGEN), Text-to-SVG Generation (TSVGEN), and SVG Understanding (SVGUN) (Li et al., 11 Aug 2025).

Task family	Training items	Benchmark items
ISVGEN	359,649	1,000
TSVGEN	90,993	500
SVGUN	75,099	1,350

The raw SVGs were collected from two open datasets, Kaggle “SVG icons” and the HuggingFace “svgen-500k-instruct”, which aggregate assets from public repositories such as SVG-Repo. After cleaning, 360k unique, renderable SVGs form the core. Each item may include SVG code, a rendered image, and a textual description called “SVGDES”, produced via GPT-4V with four parts: Overall Description, Colors, Categories, and potential real-life Usage. Task-specific question–answer prompts are added for understanding tasks such as color counting, category classification, and usage inference (Li et al., 11 Aug 2025).

The understanding component is subdivided into code-conditioned and image-conditioned variants. CSVGUN uses SVG code plus a question as input, while ISVGUN uses the rendered image plus a question. The taxonomy is explicitly tiered: easy tasks include size extraction, color extraction, shape counting, and transform counting; middle tasks include category classification and primitive descriptions; hard tasks include usage or context prediction and general description of the whole image. A larger pool of 351,350 understanding instances is constructed across categories, and the final training subset is balanced to maintain approximately a 6:1 ratio of SVGEN to SVGUN in the training split (Li et al., 11 Aug 2025).

The curation pipeline combines simple cleaning and deep cleaning. SVGs are rendered to 336×336 PNG on a white background; samples that fail to render are dropped. Deep cleaning removes DOCTYPE and XML headers, comments, redundant nested <g> groups, and superfluous spaces. Deduplication uses perceptual hashing on PNG renders, and a complexity filter discards SVGs with more than 100 Bézier curves to stabilize training workloads. The dataset is released open-source for research, while inheriting content from public upstream sources whose licenses remain relevant to redistribution (Li et al., 11 Aug 2025).

3. Representation, serialization, and benchmark metrics

UniSVG retains standard SVG grammar and shape primitives. Typical path commands include M/m, L/l, H/h, V/v, C/c, S/s, Q/q, T/t, A/a, and Z/z, alongside higher-level elements such as rect, circle, ellipse, line, polyline, and polygon. Styles include fill, stroke, stroke-width, and opacity, while transforms include translate, scale, rotate, skew, and matrix. SVG code is treated as plain text for sequence modeling, with cleaning and whitespace compression used to reduce tokenization overhead (Li et al., 11 Aug 2025).

This serialization choice has direct computational implications. UniSVG reports that removing redundant commas and decimal points reduces training tokens by 35–37% and wall-clock training time by approximately 40%, but with a small cost in generation quality. The result highlights a recurring trade-off in SVG modeling: compact textual encodings improve efficiency, while higher numeric fidelity is often required to preserve exact geometry (Li et al., 11 Aug 2025).

The official benchmark metrics are task-specific. For SVG generation, similarity is computed by rendering the generated SVG and comparing it to the reference using SSIM, LPIPS, and CLIP similarity. For SVG understanding, the benchmark uses accuracy for structured targets and BERTScore and SBERT for free-form descriptions. The public leaderboard assigns 45% of the final score to ISVGEN, 45% to TSVGEN, and 10% to SVGUN; within generation, CLIP is weighted at 60%, while SSIM and LPIPS receive 20% each. Within SVGUN, the weights for Easy-Acc, Hard-Acc, BERT, and SBERT are given as $1 : 2 : 3.5 : 3.5$ (Li et al., 11 Aug 2025).

Later work evaluating the image-to-SVG portion of UniSVG uses the benchmark’s official ISVGEN protocol and reports Low-Level, High-Level, and overall Score, following the benchmark evaluator without redefining its formulas. This indicates that UniSVG has become not merely a training corpus but an externalized evaluation substrate for code-level visual reconstruction (Zheng et al., 30 Jun 2026).

4. Empirical findings from the original benchmark

UniSVG was introduced to test whether SVG-specific finetuning can unlock MLLM capabilities for vector understanding and generation. On the benchmark, after three epochs of finetuning on UniSVG, open-source models improve substantially across all metrics, and finetuned Qwen 2.5 VL attains the best overall final score of 0.752, surpassing Claude 3.7 at 0.722, GPT-4V at 0.650, and Gemini 1.5 Pro at 0.647 (Li et al., 11 Aug 2025).

The result pattern is technically informative. Claude 3.7 achieves strong CLIP similarity, indicating high semantic alignment, but weaker SSIM and LPIPS than the finetuned open-source models. This suggests that a model can capture the semantics of an icon or graphic while still missing fine geometric detail, an especially important distinction for SVG tasks where code-level correctness is tightly coupled to rendered structure (Li et al., 11 Aug 2025).

The benchmark also supports ablation studies on training strategy. On LLaVA and LLaVA-LLaMA, finetuning the Stage-1 aligned multimodal checkpoint yields the best SVG understanding and generation performance, while additional