Resources
What is SVG training data?
Text, paths, geometry, layout and metadata: what a vector file carries that a bitmap cannot.
A file that describes itself
An SVG is a document, not a picture. Open one in a text editor and the headline is there as a string, the chart is there as coordinates, the layout is there as a hierarchy of groups. Nothing has to be recovered by inference, because nothing was ever discarded.
That is the whole argument for vector training data. A model learning from bitmaps has to reconstruct structure that the original design already had. A model learning from SVG is handed it.
What the file carries
Four kinds of signal sit in every file, each addressable on its own:
- Text as text. A text node holds the actual characters, plus the family, weight, size and position they were set in. No optical character recognition, no uncertainty about what the words are.
- Geometry as paths. Curves and lines with explicit control points, so shape is a sequence a language model can learn to write rather than a field of pixels.
- Layout as coordinates. Bounding boxes, alignment, spacing and z-order fall out of the document, which is what a layout model needs as ground truth.
- Colour as declared values. Fills, strokes and gradients are named, so the palette is recoverable exactly rather than sampled.
Why it is editable, and why that matters
Because every element can be changed and the file re-rendered, one asset can generate paired examples: move a heading and you have a layout counterfactual, change a series and you have a chart counterfactual, swap a palette and you have a style counterfactual. The pair is exact, because you produced both halves.
This is the practical difference between a corpus you can only read and a corpus you can experiment on.
What to ask of any vector corpus
Live text rather than outlined type, since outlined text is geometry pretending to be language. Per-asset metadata, so a subset can be selected rather than filtered by hand. A licence that permits commercial training. And a provenance record for each file, because a buyer who cannot say where a file came from cannot document their model.