AI training data
AI training data for structured vector graphics.
Pixels record how a design looks. A vector file records how it is made. That difference is the whole reason SVGZO exists.
What disappears when a design is flattened
Rasterising a poster keeps its appearance and destroys its structure. The headline becomes a cluster of pixels the model must OCR back into words. The chart becomes a shape with no data behind it. The layout becomes implicit. In an SVG all of that is still explicit, addressable and editable.
| Property | Raster (PNG, JPEG) | SVG |
|---|---|---|
| Text | Pixels; needs OCR; font unknown | Live <text> nodes with font family, weight, size and position |
| Shapes | Pixels | Paths with explicit geometry and Bézier control points |
| Layout | Implicit | Coordinates, groups, z-order and a viewBox |
| Colour | Sampled per pixel | Declared fills, strokes and gradients; palette recoverable |
| Scale | Fixed resolution | Resolution independent; render at any size for evaluation |
| Editability | None | Every element can be changed and re-rendered to make paired training examples |
| Size | Hundreds of KB to MB | Tens of KB per file, so a million-asset corpus is tens of gigabytes rather than terabytes |
Key concepts
- Text as content, role and typography rather than glyph pixels.
- Paths as a token sequence a language model can learn to write.
- Geometry as bounding boxes, alignment and spacing for layout models.
- Layout as page format, grid and z-order.
- Semantic structure as groups that correspond to chart, legend, title and footer.
- Editability: change one attribute, re-render, and you have a counterfactual pair.
- Metadata per asset, from sector and subject to palette and checksum.
Applications
- Text-to-SVG and image-to-SVG generation
- Native vector generation in design tools
- Layout generation and design completion
- Vector understanding and editing agents
- Multimodal learning with paired render and source
- Document and chart understanding
- Spatial reasoning over floor plans
- Evaluation of visual-generation systems against structural ground truth
Why public SVG datasets are not enough
The million-scale public corpora (SVG-Stack, MMSVG-2M, SVGX-SFT-1M, SAgoge) are icons, emoji and illustrations, and most are non-commercial. The InternSVG paper notes that existing sets pay “insufficient attention to SVGs with specific applications in professional domains.” SVGZO datasets are full pages with text, charts and layout, licensed for commercial training.
Continue to AI solutions for field-level mappings, or quality for how every file is validated.