2023
Beyond Pixels: Exploring Human-Readable SVG Generation for Simple Images with Vision Language Models
1Fudan University
2University of Illinois Urbana-Champaign
3The Hong Kong University of Science and Technology
4The Chinese University of Hong Kong, Shenzhen
Beyond Pixels is the first study to adapt a pretrained autoregressive vision-language model for image-to-SVG code generation. S²VG² turns simple raster images into human-readable SVG, with explicit shapes, colors, and spatial relationships.
Simple, readable vector graphics
The Simple-SVG-Generation method (S²VG²) targets SVG descriptions that are both visually faithful and understandable. The paper evaluates generated graphics through reasoning tasks and human judgments of readability.
Reasoning from semantic SVG
The latest manuscript includes CLEVR-SVG experiments that evaluate text-only reasoning over generated semantic SVG representations. The linked repository provides the CLEVR-SVG data and evaluation code.
Experimental results
Explore the comparisons, measurements, and examples reported in the manuscript. Each view links to its source page.
CLEVR-SVG: recorded reasoning outcomes
Two archived evaluation batches. Counts are shown separately; these are not a controlled comparison between models.
Read the scene, answer the question
Actual generated examples from the 200-scene dataset, shown with their reference answers.
Reference answer:
0
Reference answer:
magentaBibTeX
@misc{zhang2023beyondpixels,
title = {Beyond Pixels: Exploring Human-Readable SVG Generation for Simple Images with Vision Language Models},
author = {Zhang, Tong and Liu, Haoyang and Zhang, Peiyan and Cheng, Yuxuan and Wang, Haohan},
year = {2023},
eprint = {2311.15543},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2311.15543}
}