OmniSVG: A Unified Scalable Vector Graphics Generation Model

Yiying Yang1,2*, Wei Cheng2*, Sijin Chen1, Xianfang Zeng2,
Jiaxu Zhang2, Liao Wang2, Gang Yu2‡, Xinjun Ma1‡, Yu-Gang Jiang1
*Equal Contribution, ‡Correspondence Authors
1Fudan University, 2StepFun
arXiv Code Hugging Face Logo Weights Hugging Face Logo MMSVG-2M Dataset

Abstract

Scalable Vector Graphics (SVG) is an important image format widely adopted in graphic design because of their resolution independence and editability. The study of generating high-quality SVG has continuously drawn attention from both designers and researchers in the AIGC community. However, existing methods either produces unstructured outputs with huge computational cost or is limited to generating monochrome icons of over-simplified structures. To produce high-quality and complex SVG, we propose OmniSVG, a unified framework that leverages pre-trained Vision-Language Models (VLMs) for end-to-end multimodal SVG generation. By parameterizing SVG commands and coordinates into discrete tokens, OmniSVG decouples structural logic from low-level geometry for efficient training while maintaining the expressiveness of complex SVG structure. To further advance the development of SVG synthesis, we introduce MMSVG-2M, a multimodal dataset with two million richly annotated SVG assets, along with a standardized evaluation protocol for conditional SVG generation tasks. Extensive experiments show that OmniSVG outperforms existing methods and demonstrates its potential for integration into professional SVG design workflows.

Highlights

sequencing_small
OmniSVG is capable of progressively generating high-quality SVGs across a wide spectrum of complexity — from simple icons to intricate anime characters. It demonstrates remarkable versatility through multiple generation modalities, including Text-to-SVG, Image-to-SVG, and Character-Reference SVG, making it a powerful and flexible solution for diverse creative tasks.

Method Overview

sequencing_small
OmniSVG is built on a pre-trained vision-language model Qwen-VL and incorporates an SVG tokenizer. The model tokenizes both text and image inputs as prefix tokens, while the SVG tokenizer encodes vector graphics commands into a unified representation space.

Generation Process of OmniSVG

Comparison with SOTA methods

Text-to-SVG

sequencing_small

Image-to-SVG

sequencing_small

Character Reference SVG

sequencing_small

MMSVG-2M Dataset

sequencing_small
The MMSVG-2M dataset consists with three subset, namely Icon, Illustration and Character. Samples from Icon, Illustration and part of Character subsets are downloaded from Internet. Another part of Character subset is generated by our data creation pipeline, which can provide image and SVG pairs for image prompting task.

Related Links

We thank the following excellent open-source works:

IconShop: is the first advanced work that leverages LLMs to generate monochrome, icon-level SVGs. We referred to its parametric implementation.

Here is the list of highly related concurrent works:

LLM4SVG: treats SVG coordinates as number strings and predicts decimal part for higher spatial accuracy.

StarVector: equips LLM with an image encoder for Image-to-SVG generation.

Acknowledgements

We thank the following contributors for dataset construction and discussion: Haozhen Sun, Chengchen Wu, Panyi Yang, Ciba

BibTeX

@article{yang2025omnisvg,
  title={OmniSVG: A Unified Scalable Vector Graphics Generation Model},
  author={Yang, Yiying and Cheng, Wei and Chen, Sijin and Zeng, Xianfang and Zhang, Jiaxu and Wang, Liao and Yu, Gang and Ma, Xingjun and Jiang, Yu-Gang},
  journal={arXiv preprint arXiv:2504.06263},
  year={2025}
}