diffvg

Differentiable Rasterizer for Vector Graphics https://people.csail.mit.edu/tzumao/diffvg

diffvg is a differentiable rasterizer for 2D vector graphics. See the webpage for more info.

Install (Windows)

1. Install Python 3.10 from the Microsoft Store
2. Install the CUDA Toolkit
3. Install ffmpeg. Make sure ffmpeg is added to your PATH environment variable (Eg. C:\FFmpeg\bin)

> git clone https://github.com/benbaker76/diffvg
> cd diffvg
> git submodule update --init --recursive
> git -m venv .venv
> ./.venv/Scripts/activate
> pip install torch==2.4.0+cu124 torchvision==0.19.0+cu124 --index-url https://download.pytorch.org/whl/cu124
> pip install -r requirements.txt
> set LIBDIR=.\.venv\Lib\site-packages
> python setup.py bdist_wheel
> pip install dist/diffvg-0.0.1-cp310-cp310-win_amd64.whl
# Test it's working
> python apps\refine_svg.py
usage: refine_svg.py [-h] [--use_lpips_loss] [--num_iter NUM_ITER] svg target
refine_svg.py: error: the following arguments are required: svg, target

Run

cd apps

Optimizing a single circle to a target.

python single_circle.py

Finite difference comparison.

finite_difference_comp.py [-h] [--size_scale SIZE_SCALE]
                               [--clamping_factor CLAMPING_FACTOR]
                               [--use_prefiltering USE_PREFILTERING]
                               svg_file

e.g.,

python finite_difference_comp.py imgs/tiger.svg

Interactive editor

python svg_brush.py

Painterly rendering

painterly_rendering.py [-h] [--num_paths NUM_PATHS]
                       [--max_width MAX_WIDTH] [--use_lpips_loss]
                       [--num_iter NUM_ITER] [--use_blob]
                       target

e.g.,

python painterly_rendering.py imgs/fallingwater.jpg --num_paths 2048 --max_width 4.0 --use_lpips_loss

Image vectorization

python refine_svg.py [-h] [--use_lpips_loss] [--num_iter NUM_ITER] svg target

e.g.,

python refine_svg.py imgs/flower.svg imgs/flower.jpg

Seam carving

python seam_carving.py [-h] [--svg SVG] [--optim_steps OPTIM_STEPS]

e.g.,

python seam_carving.py imgs/hokusai.svg

Vector variational autoencoder & vector GAN:

For the GAN models, see apps/generative_models/train_gan.py. Generate samples from a pretrained using apps/generative_models/eval_gan.py.

For the VAE models, see apps/generative_models/mnist_vae.py.

If you use diffvg in your academic work, please cite

@article{Li:2020:DVG,
    title = {Differentiable Vector Graphics Rasterization for Editing and Learning},
    author = {Li, Tzu-Mao and Luk\'{a}\v{c}, Michal and Gharbi Micha\"{e}l and Jonathan Ragan-Kelley},
    journal = {ACM Trans. Graph. (Proc. SIGGRAPH Asia)},
    volume = {39},
    number = {6},
    pages = {193:1--193:15},
    year = {2020}
}