156×
shorter context
Vector Graphics Representation Learning
SLS: A Scalable SVG Latent Space
Learning compact, invertible path embeddings for vector graphics.
University of Modena and Reggio Emilia
Motivation
SVGs still lack the latent spaces that made raster vision scale
Raster images benefit from continuous latent spaces such as VAEs and CLIP-style embeddings, but SVGs still do not have a compact, continuous, and invertible representation. That gap matters because SVGs are symbolic, editable, and resolution-independent, which makes them attractive for graphics, design, and multimodal modeling.
Existing SVG methods typically depend on long token sequences or rasterization-heavy pipelines. SLS instead maps each SVG path to a single dense vector, giving vector graphics a representation that is compact enough for foundation-model pipelines while remaining faithful to the underlying geometry and style.
Method
From symbolic paths to invertible latent tokens
SLS encodes each SVG path into one dense latent token through a unified tokenization scheme that captures both geometry and style. The resulting embedding lies on the unit hypersphere, giving a normalized latent space that works cleanly for similarity, interpolation, retrieval, and generation.
- Path-level representation: one dense vector per SVG path instead of long XML sequences.
- Unified tokenization of geometry and style via BPE over symbolic SVG commands.
- Invertibility through a learned decoder that reconstructs the original path representation.
- Unit hypersphere embeddings for stable latent-space operations and downstream learning.
SVG path → BPE tokenizer → Transformer encoder → dense latent vector → Transformer decoder → SVG path
Key Results
Compact enough for scale, structured enough for decoding
167×
fewer training TFLOPs
Invertible
path embeddings
SLS reduces SVG captioning context length from 2432 XML tokens to 15.56 latent tokens while improving downstream quality. The result is a representation that is dramatically cheaper to model without sacrificing vector fidelity.
Reconstruction
Geometry and style are both preserved
Latent Space Robustness
Stable under meaningful perturbations
SLS remains robust under Gaussian noise and angular perturbations in latent space. This indicates that the learned representation is not only compact, but also locally well-behaved for manipulation, search, and generative modeling.
Generation
Vector-native synthesis from the learned latent space
Interpolation
A continuous manifold for SVGs
Smooth latent interpolation shows that SLS learns a structured continuous SVG manifold. Traversing between embeddings yields coherent transitions in vector geometry and style rather than brittle token-level jumps.
Applications
One latent space, multiple vector-native tasks
SVG retrieval
Compact path embeddings for similarity search across large vector collections.
SVG captioning
Short latent contexts that are much easier for language models to consume.
SVG interpolation
Continuous morphing across geometry and appearance in vector space.
Foundation models for vector graphics
A compact representation layer for multimodal and large-scale vector understanding.
Citation
BibTeX
Show citation
@inproceedings{yourpaper2026,
title={Project Title},
author={Author One and Author Two and Author Three},
booktitle={ECCV},
year={2026}
}