OmniSVG: Open-Source Multimodal SVG Generation Model by Fudan University & StepFun 60

What is OmniSVG?

OmniSVG is an open-source multimodal SVG generation model jointly released by Fudan University and StepFun. It can generate high-quality vector graphics from text, images, or character references, and was officially launched on April 9, 2025.

Technical Principles of OmniSVG

OmniSVG is built on the pretrained vision-language model (VLM) Qwen-VL and innovatively integrates an SVG tokenizer. By parameterizing SVG commands and coordinates as discrete tokens, OmniSVG successfully decouples the structural logic of vector graphics from low-level geometric details. This design not only improves training efficiency but also boosts training speed by more than 3x, while retaining the ability to generate complex SVG structures.

OmniSVG Generation Modes

• Text-to-SVG: Users can generate semantically relevant vector graphics through natural language descriptions, such as inputting "a cartoon cat sitting under a cherry blossom tree."

• Image-to-SVG: Automatically converts bitmaps (such as photos or hand-drawn sketches) into vector graphics composed of paths, preserving the visual features of the original image while gaining editability.

• Character Reference SVG: Generates vector graphics with the same character features but in different poses or scenes based on existing character images, which is particularly valuable for animation and game character design.

Datasets and Evaluation Protocol

To advance the development of SVG generation technology, the OmniSVG team released the MMSVG-2M dataset. This is a multimodal dataset containing 2 million richly annotated SVG resources, covering three major subsets: icons, illustrations, and characters. In addition, they proposed a standardized evaluation protocol, MMSVG-Bench, to test the performance of conditional SVG generation tasks.

Advantages of OmniSVG

• Compared with traditional methods, OmniSVG overcomes several core challenges of previous SVG generation techniques:

• Traditional methods often produce loosely structured, computationally expensive results, or are limited to monochrome, overly simplified icons.

• Through its end-to-end multimodal generation framework, OmniSVG significantly improves generation quality and complexity, capable of producing vector graphics that are rich in color and vivid in detail.

OmniSVG Application Scenarios

• Suitable for UI/UX design, anime character creation, industrial CAD, and other scenarios

• Generated SVGs are directly editable and support infinite zoom without loss of quality

Project Page:https://omnisvg.github.io/

Code Repository:https://github.com/OmniSVG/OmniSVG

Paper:https://arxiv.org/pdf/2504.06263