A Review of Text-Guided Vector Graphics Generation

This paper reviews recent advancements in the field of text-guided vector graphics generation, systematically categorizing current mainstream approaches into optimizationbased methods, diffusion model-based frameworks, and hybrid methods utilizing autoregressive or multimodal large models. Optimization methods rely on pre-trained models for similarity evaluation and iteratively adjust vector graphic parameters to generate results. Diffusion models leverage powerful generative capabilities to produce SVG graphics directly or indirectly, improving quality and geometric consistency. Autoregressive and multimodal methods introduce Transformers or large language models to achieve stronger structural control, better semantic alignment, and support for interactive editing of complex graphics. Although significant progress has been made in generation quality, editability, and diversity, challenges remain in terms of geometric accuracy, generation efficiency, and structural controllability. This review aims to provide comprehensive insights for researchers and developers, while highlighting key directions for future research.