This repository contains the code for the "Single Line Drawing Vectorization" paper. The full paper can be found here.

⚠️ Due to licensing issues, we are not yet able to release the full benchmark dataset we created for this paper. Samples from this dataset can be found in the sample folder.

Our code requires a Python version between 3.9 and 3.12. It is recommended to use a Python virtual environment. You can use either conda or Python's built-in venv.

After activating the virtual environment, install the base dependencies:

pip install -e .

The largest part of the code is written in pure Python. However, some components are accelerated with C++ code that needs to be compiled. The C++ parts are the voronoi diagram pruning for the medial axis computation and the Vanishing Angle computation for the medial axis pruning.

To compile these two components, simply run:

sh build.sh

Note that this command requires cmake to be installed.

Our code requires Potrace to vectorize the outline of the line drawing. The Python version of Potrace can be installed in two different ways.

PyPotrace is a Python package providing bindings to the Potrace library. This is the recommended version to install, as it provides direct bindings to the Potrace C++ library and is thus fast.

To install it, clone the repo and follow the instructions in the README.md file.

If the installation of PyPotrace failed, one can install Potracer. Potracer is a Python port of Potrace. Since everything is written in Python, the code is much slower, but also easier to install. To install it, simply run:

pip install potracer

Then switch to the potracer branch:

git checkout potracer

The weights for our intersections classification model need to be downloaded. They can be obtained here. Then the model.pth file needs to be placed in the src/SLDvec/assets/ folder.

Once everything is installed, our method can be used either with a command line interface or with the GUI. The GUI allows to interactively influence the algorithm output.

To apply our method to a single image, simply run the command:

SLDvec run IMAGE_PATH [--output-path [OUTPUT_PATH]] [--thresh [THRESHOLD]] [--multiple-lines]

--image-path            # The path to the input image containing the line drawing.
--output-path           # An optional path to save the SVG output. If not provided, the output will be saved in the same folder as the input.
--thresh                # An optional threshold value to binarize the input image. By default an automatic method is used to determine this threshold. This is the recommended method.
--multiple-lines        # A flag to use if the input image contains multiple strokes.

To apply the method to a batch of images in a folder, run:

SLDvec run-folder FOLDER_PATH [--output-dir [OUTPUT_PATH]] [--thresh [THRESHOLD]] [--multiple-lines]

--folder-path           # The path to the input folder containing the line drawings.
--output-dir            # An optional path to save the SVG outputs. If not provided, the outputs will be saved in an output folder inside the input folder.
--thresh                # An optional threshold value to binarize the input images. By default an automatic method is used to determine this threshold. This is the recommended method.
--multiple-lines        # A flag to use if the input images contain multiple strokes.

The GUI is a web app that can be launched using

SLDvec gui [--port PORT] 

--port                # The port on which to launch the web app, by default 5000.

The web app will be launched at http://127.0.0.1:5000/ by default. Information on how to use the GUI can be found at the end of the supplementary video.

If you used this code, please consider citing our work:

@article{Magne:SLDvectorization:2025,
title = {Single-Line Drawing Vectorization},
author = {Tanguy Magme and Olga Sorkine-Hornung},
journal = {Computer Graphics Forum (proceedings of Pacific Graphics 2025)},
volume = {44},
number = {7},
year = {2025},
url = {https://doi.org/10.1111/cgf.70228},
}