SLS — A Scalable Vector Graphics Latent Space

Pretrained weights and tokenizer for A Scalable Vector Graphics Latent Space (ECCV 2026).

• Authors: Leonardo Zini, Elia Frigieri, Lorenzo Baraldi
• Paper: https://arxiv.org/abs/2608.21893
• Code: https://github.com/aimagelab/svg_latent_space
• Project page: https://aimagelab.github.io/svg_latent_space/

Model

SLS is a Transformer autoencoder that maps each SVG path to a 1024 dimensional latent vector and decodes it back to symbolic geometry and style. The released model has 8 encoder layers, 8 decoder layers, 8 attention heads, a vocabulary of 448 tokens, and 135,774,656 learned parameters.

This is a custom PyTorch model loaded using the code repository linked above. The architecture is described in config.json, and configs/sls.yaml contains the training and inference settings used by the code.

Files

weights/sls/model.safetensors
tokenizer/tokenizer.model
tokenizer/tokenizer.vocab
configs/sls.yaml
config.json
assets_manifest.json
LICENSE

The release contains the pretrained model and tokenizer. Optimizer, scheduler, RNG states, training logs and datasets are not included. SHA256 checksums for the model and tokenizer are recorded in assets_manifest.json.

Quick start

Clone the code repository and install its dependencies using Python 3.11:

git clone https://github.com/aimagelab/svg_latent_space.git
cd svg_latent_space
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Download the assets and set the required environment variable:

hf download lzini/SLS --local-dir sls_release
export SLS_ASSETS_DIR="$(realpath sls_release)"

In this version, SLS_ASSETS_DIR must be exported before using the default config for training or inference. Repeat the export in each new terminal.

python inference.py reconstruct --input examples/example.svg --output outputs/reconstructed.svg
python inference.py encode --input examples/example.svg --output outputs/embeddings.safetensors
python inference.py decode --input outputs/embeddings.safetensors --output outputs/decoded.svg

CUDA is used when available. Add --device cpu to select CPU execution. See the code repository for training, fine-tuning and checkpoint-resume commands.

Representation and limits

The model uses last-token pooling (cls_mode: last) and produces raw latent vectors (l2norm: false). Preserve their magnitude when decoding. Normalize a separate copy when computing cosine similarity.

The inference code canonicalizes supported SVG shapes into paths on a 512×512 canvas, with rounded coordinates and serialized style. Inputs exceeding 1024 tokens per path are rejected. Arbitrary CSS and nested transforms are not fully supported, and predicted SVG/XML paths may require validation.

License

The released weights and tokenizer are licensed under Apache-2.0. See LICENSE.

Citation

@inproceedings{zini2026scalable,
  title     = {A Scalable Vector Graphics Latent Space},
  author    = {Zini, Leonardo and Frigieri, Elia and Baraldi, Lorenzo},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}