ReDualSVG: Refined Scalable Vector Graphics Generation

Shannan Yan

Abstract: Vector graphics generation is a critical task in computer vision. However, existing approaches suffer from several limitations, such as lack of extensibility, inadequate consideration of both image and sequence modalities, and the issue of location change. To address these challenges, we present ReDualSVG, a refined scalable vector graphics generation method based on dual-modality information. ReDualSVG overcomes these problems through a hierarchical, transformer-based design that effectively distinguishes high-level shapes from the low-level instructions encoding the shape itself. Our method capitalizes on dual-modality data by fully utilizing the information present in both image and sequence modalities, offering richer insights from global and local perspectives. Additionally, we tackle the issue of location change by employing a differentiable rasterizer for further refinement. Finally, we conducted qualitative and quantitative experiments on a publicly available dataset, which demonstrated that ReDualSVG achieves high-quality synthesis results in the applications of image reconstruction and interpolation, outperforming other alternatives.