AutoFigure-Edit transforms paper method sections into fully editable SVG figures and allows for refinement in an embedded SVG editor. It addresses the need for publication-ready figures that can be easily modified without relying on raster images. The core problem is the difficulty of editing figures generated from text descriptions using traditional methods. This project leverages a combination of large language models and semantic segmentation to generate editable vector graphics.
AutoFigure-Edit distinguishes itself by producing fully editable SVG figures, unlike raster images. The inclusion of an embedded SVG editor facilitates direct refinement, enabling users to adjust text, shapes, and layout. Style transfer capabilities allow for mimicking artistic styles found in reference images. The system's use of SAM3 for icon detection jointly with a flexible template makes it exceptionally robust to vary figure structures.
- Text-to-Figure Generation: Converts method text into draft figures.
- SAM3 Icon Detection: Accurately detects and segments icons from generated figures.
- Editable SVG Output: Produces fully editable SVG files for easy modification.
- In-Browser Editor: Offers a convenient embedded SVG editor for refinement.
- Style Transfer: Implements style transfer based on provided reference images.
- Automated Templating: Generates SVG templates aligned with figure structure.
- Artifact Generation: Generates PNG/SVG outputs and icon crops.
AutoFigure-Edit is a relatively new project with active development. The online platform is available for public use, demonstrating functional capabilities. Recent activity includes paper acceptance to ICLR 2026, suggesting ongoing research and refinement. Documentation is being developed, and a growing community is forming around the project through discussion on Hugging Face.
Researchers, scientists, and anyone needing to create and modify publication-quality figures will benefit from AutoFigure-Edit. Its ability to transform text into editable SVGs streamlines the figure creation process, allowing for easy customization and ensuring consistency across publications. It offers a significant value proposition over manual vectorization or using raster images, providing a structured and easily adaptable solution.