AB
AB
AiBoss

All Categories

project

Custom-SVG - An SVG customization framework jointly developed by Adobe and City University of Hong Kong.

Custom-SVG is a two-stage SVG customization framework developed by Adobe and City University of Hong Kong. It generates high-quality SVG graphics with customized styles based on text prompts. The framework introduces a two-stage generation workflow...

What is Custom-SVG?

Custom-SVG is a two-stage style-customizing SVG generation framework developed by Adobe and City University of Hong Kong. It generates high-quality SVG graphics with customized styles based on text prompts. The framework introduces a two-stage generation workflow, learning the structure and content of SVG based on a path-level representation diffusion model. It extracts style information from a customized text-to-image (T2I) model, enabling diverse style customization. Custom-SVG combines the efficiency of a feedforward model with the powerful generative capabilities of a diffusion model to quickly generate well-structured and stylistically consistent vector graphics, suitable for design, icon generation, and other scenarios.

Main functions of Custom-SVG

  • Maintaining the structural regularity of SVGThe generated vector graphics have a clear hierarchical structure and regular paths, making them easy to edit later.
  • Achieve style customizationGenerates diverse SVGs that match a small number of style examples provided by the user.
  • High-efficiency generationBased on a feedforward generation method, it generates results quickly and is suitable for practical design scenarios.
  • Semantic alignmentThe generated SVG content is highly consistent with the input text prompts, ensuring the semantic accuracy of the graphics.

The technical principles of Custom-SVG

  • Training of T2V diffusion model based on path-level representation:
    • Path-level representation: Decompose SVG into path-level representations, each path is defined by parameters such as Bézier curves and colors, and encoded into compact latent space vectors based on pre-trained SVG VAEs.
    • diffusion modelThe diffusion model is used to gradually recover the SVG tensor aligned with the text prompt from the noise.
    • Transformer architectureUsing Transformer as the backbone network, combined with self-attention and cross-attention mechanisms, the model can understand the semantics of the text and generate the corresponding SVG structure.
  • based onImage diffusion prior for style customization:
    • Style ExtractionFine-tune the T2I diffusion model to generate diverse custom images using a small number of style examples.
    • Image-level lossThe generated SVG is rendered as an image, and the T2V model is fine-tuned based on an image-level loss function to give the generated SVG the desired style.
    • Style transferThis allows for flexible style customization by transferring the style of custom images to the SVG generation process.

Project address for Custom-SVG

Application scenarios of Custom-SVG

  • Graphic Design and IllustrationQuickly generate vector graphics that conform to a specific style for design concept validation, illustration creation, and brand design.
  • User Interface (UI) DesignGenerates interface elements such as icons and buttons to meet the needs of different interface designs.
  • Web DesignGenerate dynamic vector graphics to enhance the visual appeal and responsive design of web pages.
  • Education and TrainingGenerate teaching graphics and interactive learning tools to assist in the presentation of teaching content.
  • Personalized content generationGenerate personalized graphics based on user needs, which can be used in customized gifts, social media content, etc.
Back