SVGDreamer++: Advancing Editability and Diversity in Text-Guided SVG Generation

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Text-guided scalable vector graphics (SVG) synthesis is promising for iconography and sketching.
• Existing Text-to-SVG methods struggle with editability, visual quality, and diversity.

Purpose of the Study:

• To develop a novel framework for text-guided SVG synthesis that enhances editability, quality, and diversity.
• To enable fine-grained editing and improve the aesthetic appeal and detail presentation of generated SVGs.

Main Methods:

• Introduced Hierarchical Image VEctorization (HIVE) for object-level semantic control and component optimization, enhancing editability.
• Developed Vectorized Particle-based Score Distillation (VPSD) to address over-saturation and boost sample diversity, incorporating a reward model for aesthetic improvement.
• Designed an adaptive vector primitives control strategy for dynamic adjustment of graphic details.

Main Results:

• The proposed framework significantly improves editability, visual quality, and diversity compared to baseline methods.
• Demonstrated superior performance in generating precise vector graphics with fine-grained editing capabilities.
• Achieved enhanced sample diversity and aesthetic appeal through VPSD and adaptive primitive control.

Conclusions:

• The novel framework offers a superior approach to text-guided SVG synthesis, overcoming limitations of existing methods.
• The method supports multiple distinct vector styles, enabling high-quality vector asset generation for stylized design and posters.
• The advancements in editability, quality, and diversity pave the way for more sophisticated applications of SVG synthesis.