Custom-SVG - An SVG custom generation framework launched by Adobe and City University of Hong Kong

What is Custom-SVG

Custom-SVG is a two-stage style SVG custom generation framework launched by Adobe and City University of Hong Kong, generating high-quality SVG graphics with custom styles based on text prompts. The framework introduces a two-stage generation process, learning the structure and content of SVG based on a diffusion model with path-level representation, extracting style information from a customized text-to-image (T2I) model, and achieving diverse style customization. Custom-SVG combines the efficiency of feedforward models with the powerful generation capability of diffusion models, quickly generating vector graphics with regular structure and consistent style, suitable for scenarios such as design and icon generation.

Main Features of Custom-SVG

• Maintaining SVG Structural Regularity: The generated vector graphics have a clear hierarchical structure and regular paths, facilitating subsequent editing.
• Achieving Style Customization: Based on a small number of style examples provided by the user, generate SVG in diverse styles that match them.
• Efficient Generation: Based on a feedforward generation method, quickly generate results, suitable for actual design scenarios.
• Semantic Alignment: The generated SVG content is highly consistent with the input text prompt, ensuring the semantic accuracy of the graphics.

Technical Principles of Custom-SVG

• T2V Diffusion Model Training Based on Path-Level Representation：
  • Path-Level Representation: Decompose SVG into path-level representations, where each path is defined by parameters such as Bézier curves and colors, and is encoded into compact latent space vectors based on a pretrained SVG VAE.
  • Diffusion Model: Use the denoising process of the diffusion model to gradually recover SVG tensors aligned with the text prompt from noise.
  • Transformer Architecture: Use Transformer as the backbone network, combining self-attention and cross-attention mechanisms, enabling the model to understand text semantics and generate the corresponding SVG structure.
• Style Customization Based onImage Diffusion Priors：
  • Style Extraction: Fine-tune the T2I diffusion model, using a small number of style examples to generate diverse customized images.
  • Image-Level Loss: Render the generated SVG as images, and fine-tune the T2V model based on an image-level loss function, so that the generated SVG has the desired style.
  • Style Transfer: Transfer the style of customized images into the SVG generation process, achieving flexible style customization.

Custom-SVG Project Address

• Project Website：https://customsvg.github.io/
• GitHub Repository：https://github.com/intchous/custom-svg-style
• arXiv Technical Paper：https://arxiv.org/pdf/2505.10558

Application Scenarios of Custom-SVG

• Graphic Design and Illustration: Quickly generate vector graphics that conform to a specific style, for design concept validation, illustration creation, and brand design.
• User Interface (UI) Design: Generate interface elements such as icons and buttons, meeting the needs of different interface designs.
• Web Design: Generate dynamic vector graphics, enhancing web visual effects and responsive design.
• Education and Training: Generate instructional graphics and interactive learning tools, assisting the presentation of teaching content.
• Personalized Content Generation: Generate personalized graphics according to user needs, used for customized gifts, social media content, etc.