NeuralSVG: Text-to-Vector Graphics with Layered and Editable SVGs

NeuralSVG: An Implicit Representation for Text-to-Vector Generation

ICCV 2025

Authors:

• Sagi Polaczek (Tel Aviv University)
• Yuval Alaluf (Tel Aviv University)
• Elad Richardson (Tel Aviv University)
• Yael Vinker (Tel Aviv University, MIT CSAIL)
• Daniel Cohen-Or (Tel Aviv University)

arXiv | Code

Overview

Vector graphics are a cornerstone in design, offering resolution-independent and highly editable visual content. Recent advancements in vision-language models and diffusion models have sparked interest in generating vector graphics from text prompts. However, existing approaches often produce over-parameterized outputs and treat the layered structure of SVGs as a secondary goal, limiting their practicality.

NeuralSVG addresses these issues by introducing an implicit neural representation for generating vector graphics from text. Inspired by Neural Radiance Fields (NeRFs), NeuralSVG encodes the entire scene into the weights of a small Multi-Layer Perceptron (MLP) network, optimized using Score Distillation Sampling (SDS). This approach not only ensures a structured and editable SVG but also allows for dynamic conditioning, such as changing background colors or aspect ratios.

How It Works

• Implicit Neural Representation: NeuralSVG learns an implicit representation of the vector graphics by encoding the SVG into the weights of a small MLP network. The model is optimized using Score Distillation Sampling (SDS), which helps in generating high-quality and structured SVGs.

• Dropout-Based Regularization: To promote an ordered and meaningful layered structure, NeuralSVG uses a dropout-based regularization technique. This method encourages each shape to have a standalone meaning within the overall scene, ensuring that the generated SVG is both structured and editable.

• Inference-Time Control: The neural representation provides inference-time control, allowing users to dynamically adjust aspects of the generated SVG, such as color palettes or aspect ratios, using a single learned representation. This flexibility is crucial for practical applications in design and graphics.

Technical Details

• Model Architecture:

  • NeuralSVG uses a small MLP network to encode the entire scene.
  • The network is optimized using Score Distillation Sampling (SDS), which helps in generating high-quality and structured SVGs.
  • Dropout-based regularization is applied during training to encourage each shape to have a meaningful role in the overall scene.

• Dynamic Conditioning:

  • NeuralSVG supports dynamic conditioning, allowing for changes in background color, aspect ratio, and other properties at inference time.
  • This feature enables users to generate multiple variations of the same SVG with different color palettes or layouts using a single learned representation.

Examples

NeuralSVG can generate vector graphics from text prompts while maintaining a structured and editable format. For example:

• Text Prompt: "A red apple on a white background"

  • NeuralSVG generates an SVG of a red apple, with each shape (e.g., the apple, stem, and shadow) having a meaningful role in the overall scene.
  • Users can dynamically change the background color or adjust the aspect ratio without retraining the model.

• Text Prompt: "A blue car on a yellow road"

  • NeuralSVG produces an SVG of a blue car on a yellow road, with each component (e.g., the car body, wheels, and road) being editable.
  • Users can experiment with different color palettes or adjust the layout to fit their design needs.

Evaluation

NeuralSVG has been evaluated both qualitatively and quantitatively. The results show that it outperforms existing methods in generating structured and flexible SVGs. Key findings include:

• Qualitative Results: Generated SVGs are highly structured and editable, with each shape having a meaningful role in the overall scene.
• Quantitative Results: NeuralSVG achieves higher scores in terms of structural accuracy and editability compared to other text-to-vector generation methods.

Conclusion

NeuralSVG represents a significant step forward in text-to-vector graphics generation. By leveraging an implicit neural representation and dropout-based regularization, it ensures that generated SVGs are structured, editable, and flexible. The ability