What happens when you ask large language models to recreate famous paintings as simple vector graphics? This gallery shows their attempts to turn artworks into SVG images. What makes this interesting isn’t that GenAI can generate images from prompts - specialized models already do that. It’s how they combine knowledge of art with the ability to write SVG code.

The project is inspired by Simon Willison's practice of using this as a creative LLM benchmark. When a new model is released, one of his first tests is whether it can generate a pelican on a bicycle.

It also comes with a workshop feature that I use to add or generate new artworks. You can use it yourself as well by starting the project locally (and providing an OpenRouter API key). You can see it in action here: https://x.com/vnglst/status/1971461976040800730