Single-Line Drawing Generation via Semantics-Driven Optimization

Tanguy Magne, Alexandre Binninger, Ruben Wiersma, Olga Sorkine-Hornung

This repository contains the code for the CGF paper "Single-Line Drawing Generation via Semantics-Driven Optimization".

The code is partially based on the ControlSketch part of SwiftSketch. We thank the authors for sharing their work.

🐋 Docker

A prebuilt Docker image is available on Docker Hub so you can reproduce the results without going through the installation steps below. See the tanguymagne/sldgen repository on Docker Hub for prerequisites, the run command, and how to mount your data and Hugging Face cache.

🛠️ Installation

1. Create and set up a new conda environment

conda create -n sldgen python=3.9.19 -y
conda activate sldgen
conda install cuda -c nvidia/label/cuda-12.4.0
pip install "setuptools<78" wheel

2. Install diffvg

This is the differentiable rasterizer used to optimize the curve parameters. The following instructions are taken from their installation guide.

git clone git@github.com:BachiLi/diffvg.git
cd diffvg
git submodule update --init --recursive
conda install -y pytorch torchvision -c pytorch
conda install -y numpy
conda install -y scikit-image
conda install -c conda-forge cmake=3.28
conda install -y -c conda-forge ffmpeg
pip install svgwrite
pip install svgpathtools
pip install cssutils
pip install numba
pip install visdom --no-build-isolation
pip install torch-tools
python setup.py install

3. Install the repulsion loss

The repulsion loss uses the implementation from the Fabricable 3D Wire Art paper. To install:

git clone git@github.com:kenji-tojo/fab3dwire.git
cd fab3dwire && cd wiregrad
conda install conda-forge::eigen
pip3 install --no-build-isolation git+https://github.com/openai/CLIP.git@a1d071733d7111c9c014f024669f959182114e33
pip3 install -r ./requirements.txt
pip3 install .

4. Install the TSP solver

We use the Concorde TSP solver to initialize the curve before optimization. Full installation instructions are available here.

In short, first download the source code for version 03.12.19 from here, then extract it:

wget https://www.math.uwaterloo.ca/tsp/concorde/downloads/codes/src/co031219.tgz
gunzip co031219.tgz
tar xvf co031219.tar
cd concorde

Download qsopt.a and qsopt.h from here, and place them in the same folder (note that the following links are for Ubuntu):

wget https://www.math.uwaterloo.ca/~bico/qsopt/downloads/codes/ubuntu/qsopt.a
wget https://www.math.uwaterloo.ca/~bico/qsopt/downloads/codes/ubuntu/qsopt.h

Then run:

./configure --with-qsopt=/full/path/to/concorde/folder --enable-ccdefaults
make

It is important to supply the full path to the Concorde folder. Relative paths or paths that use ~ will not work.

This should create a TSP folder containing a concorde binary. Create an environment variable CONCORDE_PATH pointing to this binary:

export CONCORDE_PATH='[path to concorde]/TSP/concorde'

5. Install other dependencies

pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txt

6. Authenticate with Hugging Face

Follow these instructions to create a Hugging Face user access token and authenticate.

Then request access to Stable Diffusion 3.5 medium using the same account.

▶️ Running

Running the code requires a machine with a GPU with at least 24GB of VRAM (tested on an RTX 4090).

To run the code and generate a single-line drawing of a given image, simply run:

python sldgen.py --target ./data/firefighter.png

This command uses default parameters that often give the best results. However, many parameters can be tuned. For more info, check the docs or run:

python sldgen.py --help

🪪 Citation

@article{Magne:SLDgen:2026,
    title   = {Single Line Drawing Generation via Semantics-Driven Optimization},
    author  = {Magne, Tanguy and Binninger, Alexandre and Wiersma, Ruben and Sorkine-Hornung, Olga},
    journal = {Computer Graphics Forum},
    volume  = {n/a},
    number  = {n/a},
    pages   = {e70502},
    year    = {2026},
    doi     = {https://doi.org/10.1111/cgf.70502},
}