Jan 09, 2025 15:00:00

MIT research team announces 'NeuralSVG', an AI tool that can generate vector images composed of multiple shapes from text

A research team from

Tel Aviv University and the Massachusetts Institute of Technology's Computer Science and Artificial Intelligence Laboratory has announced NeuralSVG , an AI tool that generates vector images consisting of multiple shapes based on text prompts.

[2501.03992] NeuralSVG: An Implicit Representation for Text-to-Vector Generation
https://arxiv.org/abs/2501.03992

NeuralSVG
https://sagipolaczek.github.io/NeuralSVG/

Vector images are images that are expressed as a collection of shapes such as circles and straight lines, and are widely used in the design field because they allow for advanced editing of size and shape. With the recent development of visual language models and diffusion models , interest in generating vector images based on text prompts is also growing.

However, existing approaches to generating vector images from text often have poor practical use because the output is overly parameterized or the layer structure of shapes, an important element of vector images, is not taken into consideration.

Therefore, the research team developed an AI tool called 'NeuralSVG' that generates vector images from text with emphasis on layer structure. This tool generates practical vector images by weighting vectors with a technique called Score Distillation Sampling (SDS) and encouraging each shape to be layered using a dropout- based method.

The vector images generated by NeuralSVG are composed of various shapes as shown below.

Vector images generated by NeuralSVG can capture the rough overall structure even with a small number of shapes, and allow for advanced editing according to layers.

Users can also specify a background color for the image, and NeuralSVG generates vector images with different color palettes depending on the specified color. The figure below shows the color palettes observed during training (top) and in general use (bottom).

In NeuralSVG, if you specify the aspect ratio of a vector image, the corresponding image will be generated. In the following figure, the image on the left is generated with an aspect ratio of 1:1, and the image on the right is generated with an aspect ratio of 4:1.

Additionally, NeuralSVG can generate sketch-like vector images with a specified number of lines without modifying the entire framework.

The NeuralSVG code will be made publicly available on GitHub in the near future.

GitHub - SagiPolaczek/NeuralSVG: Official implementation of NeuralSVG
https://github.com/SagiPolaczek/NeuralSVG