Summary

OmniSVG is a unified framework for generating high-quality Scalable Vector Graphics (SVG) through a multimodal approach, developed by Fudan University and the StepFun team and presented at NeurIPS 2025. It leverages advanced pre-trained vision-language models, specifically Qwen2.5-VL, to convert text descriptions, raster images or character references into fully editable vector graphics. Unlike traditional tools that often produce unstructured or oversimplified SVGs, OmniSVG excels at creating complex, visually appealing and logically organized vector images, ranging from simple icons to intricate illustrations and character designs.

Key Features

Text-to-SVG generation for converting descriptions into editable vector graphics
Image-to-SVG conversion for transforming raster images into structured SVGs
Character-reference SVG generation for style-consistent illustrations
Supports highly detailed and complex SVG structures with over 30,000-token contexts
Built on the Qwen2.5-VL vision-language model with a dedicated SVG tokenizer
MMSVG-2M dataset of two million richly annotated SVG assets
Standardized MMSVG-Bench evaluation protocol for conditional SVG generation
Open-source code, models and datasets for research and development

About OmniSVG


A core strength of OmniSVG lies in its technical architecture, which transforms SVG commands and coordinates into discrete tokens, separating structural logic from geometric details, with paths standardized into a small vocabulary of commands plus fill-color tokens. This tokenization enables efficient training and expressive SVG generation, supporting long context sequences of over 30,000 tokens for highly detailed outputs. The system employs a Vision Transformer-based image encoder and a dedicated adapter to project image embeddings into the latent space of the large language model, facilitating seamless multimodal input handling.


OmniSVG also supports character-reference SVG generation, enabling users to create consistent vector illustrations of specific characters in various poses and scenarios, which is particularly valuable for animation, branding and game design. The project is open source, with the MMSVG-2M dataset of two million richly annotated SVG assets and the MMSVG-Bench evaluation protocol, and it outperforms prior methods on text-to-SVG and image-to-SVG benchmarks.

Get more likes & reach the top of search results by adding this button on your site!

Embed button preview - Light theme
Embed button preview - Dark theme
TurboType Banner

Subscribe to the AI Search Newsletter

Get top updates in AI to your inbox every weekend. It's free!