Show HN: Browse how LLMs generate famous artworks as SVGs (pelican.koenvangilst.nl)

A new gallery collects examples of large language models asked to recreate famous paintings as SVG vector images, not raster pictures. Rather than using dedicated image generators, the project prompts LLMs to output executable SVG code (paths, shapes, fills, strokes, transforms) that approximates well‑known artworks. The novelty lies in combining art knowledge with programmatic code generation: the results are compact, inspectable, and immediately renderable, exposing how models translate semantics like “cubes in perspective” or “muted ochre palette” into concrete SVG primitives. This is significant for the AI/ML community because it serves as a lightweight, interpretable benchmark of spatial reasoning, compositional understanding, and symbolic output fidelity. Unlike pixel generators, SVG outputs reveal where models grasp structure (correct layering, transforms, coordinate systems) and where they fail (misplaced coordinates, incorrect color codes, broken path syntax). Inspired by Simon Willison’s informal “pelican on a bicycle” test, the gallery provides a reproducible way to stress test new models’ creativity and programmatic control. Practical implications include automated vector art tools, programmatic icon generation, and a diagnostic lens for prompt engineering and model capability evaluation.