Generating SVG Images from Text Descriptions

Objective

Develop a model that generates Scalable Vector Graphics (SVG) code based on a given text prompt, rendering the described image as accurately as possible.

Problem Statement

Large Language Models (LLMs) often struggle to generate precise image-rendering code. This project aims to bridge that gap by generating SVG, a vector image format that uses XML to define two-dimensional graphics. SVGs can be scaled without quality loss, making them ideal for various applications.

Dataset

The dataset consists of 500 text descriptions of everyday objects and scenes across diverse domains.

Description Properties

• Common objects and generic subjects (no brand names, trademarks, or personal names).
• Covers approximately a dozen categories, including landscapes, abstract art, and fashion.
• Descriptions are capped at 200 characters, with an average length of 50 characters.
• train.csv includes data from the landscape, abstract, and fashion categories.
• Public and private test sets follow a similar category distribution.

Submission Requirements

Participants must submit a Model class with a predict() function that takes a text description as input and returns SVG code.

Evaluation Criteria

Performance is measured using Mean CLIP Similarity between the text description and the generated SVG image:

1. Each SVG is converted into a PNG using the cairosvg Python library.
2. The PNG image is encoded into feature embeddings using a SigLIP SoVIT-400m model.
3. The final score is the average cosine similarity between the text description embeddings and the corresponding image embeddings.

Constraints

SVG Constraints

• The SVG file must not exceed 10,000 bytes.
• Only elements and attributes from a predefined allowlist are permitted.
• CSS styles are not allowed.
• No rasterized images or external data sources are allowed in the SVG.

Model Constraints

• The model must return SVG output within 5 minutes of receiving a description.
• All SVGs must be generated within 9 hours.

Installation & Usage

Dependencies

To run the model, install the required dependencies:

pip install cairosvg torch torchvision

Running the Model

Ensure you have the dataset and then run the model:

from model import Model

model = Model()
text_description = "A simple mountain landscape with a rising sun."
svg_output = model.predict(text_description)
print(svg_output)  # Outputs SVG code