How They Are Made
No language model writes these SVGs. The pipeline is raster-first: make a great picture, turn it into honest vectors with real tools, and check the result twice. Also, a human watches.
1. Generate
We ask an image model for a pelican riding a bicycle, in compositions built to vectorize: crisp edges, no text, no watermarks, a mechanically coherent bike, an unmistakable pouch. Budget discipline applies — roughly two generations per accepted pelican — and new sources come out square, 1:1, like a proper print.
2. Screen
Every candidate is screened by vision before it earns a trace: anatomy, bicycle mechanics, bird–bicycle interaction, composition. Studio-style photoreal keeps failing review; environmental photoreal, flat editorial, and poster styles pass. The pipeline has opinions now.
3. Preprocess
ImageMagick resizes and quantizes. rembg removes backgrounds when a clean edge is worth the effort. Whatever the source needs to trace cleanly, it gets.
4. Vectorize
VTracer sweeps parameters — palette sizes, color precision, path precision, filtering — and we keep the best candidates. Each approved image produces two editions: the Visual Master (maximum fidelity, all the shading, up to tens of thousands of paths) and the Teaching Vector (fewer colors, fewer paths, tidy order-preserving groups, built to be useful).
5. Render & compare
Every SVG renders twice — resvg and librsvg — because a pelican that renders in one tool is a rumor, not a fact. Fidelity against the source is scored deterministically, with no vibes allowed.
6. Review
A human looks at each pair and votes pass or fail. Failures get reworked once — we believe in second chances, not fourth ones — and a second failure retires the item with full honors. One pilot candidate is already retired.
7. Publish
Approved items ship into the dataset with prompts, captions, hashes, settings, and provenance. Everything is CC0. Nothing is hidden.
8. The standards
- Standalone SVG only: no embedded rasters, no scripts, no external requests, no fonts.
- Identical renders in resvg and librsvg before anything ships.
- Every file hashed; if a file changes, the published hash catches it.
Benchmark background and contamination caveats: docs/benchmark.md · Licensing: docs/licensing.md.