About

The project

Better Pelicans is an open collection of extraordinarily high-quality SVG illustrations of pelicans riding bicycles. The collection is produced by a raster-first pipeline, reviewed by humans, and published openly under CC0 with full provenance.

Where the idea came from

In October 2024, Simon Willison asked AI models to generate an SVG of a pelican riding a bicycle, explained why that is harder than it sounds — distinct objects, plausible riding, depth, a valid standalone SVG — and published the benchmark. He inspired this project, but he did not create or run it and is not affiliated with it. We took his idea extremely seriously — possibly too seriously.

Why pelicans

It is the best test of compositional image generation we know: two distinct objects, one absurd premise, and a dozen ways to fail. If the pipeline works on a bird on a bike, it works on serious subjects. Also, pelicans are excellent.

How the project runs

Staged experiments with review gates, recorded budgets, and a coffee-money operating model. No bulk generation without sign-off; every accepted pelican costs about two image generations, a pile of CPU, and one very careful human review.

Where things stand

• Pilot release pilot-0003: four approved pelicans, eight SVGs, CC0 — on preview in this site's gallery.
• Public release: held until the collection clears the review bar. The pelican is patient.
• Infrastructure: site, pipeline, validation, and the review process are all operational.

Provenance & licensing

Every artifact carries an honest account of how it was made. Sources are generated with gpt-image-2-medium via OpenAI/Codex; traces are made with VTracer 0.6.12; every SVG renders under resvg and librsvg before shipping; every file is SHA-256-hashed. No claim of human authorship is made for model-generated content. Assets are CC0-1.0; code is MIT. Full policy: docs/provenance.md · docs/licensing.md.

Contact

[email protected] · Security issues: SECURITY.md.