SVG Basic Benchmark Dataset (svg_basic_benchmark_v0)

Presto Design is proud to announce this public benchmark to help advance the field of machine-powered graphic design.

Why This Matters

Professional graphic design is increasingly reliant on Large Language Models (LLMs) due to their unique capabilities that traditional image generation models like Stable Diffusion cannot match. LLMs excel at:

• Working with branded assets and specific stock photos
• Utilizing brand-specific fonts and typography
• Creating scalable, resolution-independent designs
• Maintaining precise control over design elements
• Generating semantic, editable markup (SVG)

As a leader in developing cutting-edge LLM-powered graphic design models, Presto Design recognizes that LLMs have historically struggled with graphic design tasks. This benchmark represents a "fifth grader test" for LLMs - given an image, can they perfectly replicate it?

This fundamental capability is a crucial stepping stone toward developing truly creative and professional design systems.

Dataset Contents

• 2000 test samples, each containing:
  • image: A rendered PNG version of the SVG poster (280x280 pixels for speed and efficiency)
  • completion: The full SVG markup for the poster

Design Features

The benchmark tests comprehensive understanding of SVG features including:

• Backgrounds: Plain colors, gradients, or background images with controlled opacity
• Layouts: Text wrapping and random layouts with weighted distribution
• Text: Various font styles, sizes, and colors using a consistent color scheme
• Shapes: Geometric shapes and masked images
• Images: Integration of stock photos from the Unsplash dataset with proper attribution
• Advanced Features: masks, gradients, strokes, icons
• Color Replication

Technical Details

• SVG files are fully valid and renderable
• Images are resized to maintain performance
• Color schemes are programmatically generated for consistency
• All external resources (images, fonts) are properly embedded

Usage

This dataset is primarily used for:

1. Benchmarking SVG generation models
2. Testing SVG manipulation and rendering capabilities
3. Evaluating layout algorithm performance
4. Validating text placement and wrapping functionality

Resources

• Dataset: https://huggingface.co/datasets/Presto-Design/svg_basic_benchmark_v0
• Benchmark Runner: https://github.com/Presto-design/svg-benchmark

Dataset Structure

The dataset is organized with the following structure:

• test/: Contains 2000 examples across 1 chunks

Loading the Dataset

You can load this dataset using the Hugging Face datasets library:

from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("Presto-Design/svg_basic_benchmark_v0")

# Or load specific splits
dataset = load_dataset("Presto-Design/svg_basic_benchmark_v0", split="test")