Drawing with LLMs

May 2025 - recently there was a competition called Drawing with LLMs, where the goal was to create SVG code representing descriptions of everyday scenes in less than 10kb. The best approach overall turned out to be hacking the competition metric, while the best image generation approach was to generate diffusion model image outputs and convert them to SVGs. Actual LLMs were far down on the tier list.

This page shows examples of the pinnacle of SVG art, as generated by LLMs. The current ones below are from the OpenAI o3 reasoning model.

Landscapes

Animals & Nature

Abstract & Geometric