Yash Belhe, Ishit Mehta, Wesley Chang, Iliyan Georgiev, Michaël Gharbi, Ravi Ramamoorthi, Tzu-Mao Li
SIGGRAPH Asia 2025 (Journal Track) - Best Paper Award!

Project Website · Paper PDF · Video

We introduce automatic boundary sampling for discontinuities in differentiable shaders. Rather than writing specialized boundary queries for every primitive, our method transforms shaders into surrogate, piecewise-constant programs and samples discontinuities by snapping segments to discontinuities. This repository contains the Slang/PyTorch re-implementation for the experiments in the paper.

• Linux with CUDA-capable GPU (tested with CUDA 12.x); not tested on other operating systems.
• Conda or Miniforge
• A recent NVIDIA driver that matches your CUDA toolkit

git clone https://github.com/yashbelhe/auto_disc_sample.git
cd auto_disc_sample
conda env create -f environment.yml
conda activate asd
export CUDA_HOME=/usr/local/cuda      # adjust if CUDA is installed elsewhere
export PATH="$CUDA_HOME/bin:$PATH"

The environment installs PyTorch, SlangTorch, PyKeOps (for KDE-based boundary sampling), and the visualization stack (Matplotlib, Pillow, ImageIO). If you prefer a different CUDA toolkit, adjust the versions inside environment.yml accordingly.

Examples compile the required Slang shaders on first run via compiler/compile_shader.py. The generated files (with __gen__ prefixes) are stored under slang/ and re-used on subsequent runs. Delete them if you need to force recompilation after modifying a shader.

Note: The compiler has been recently refactored and my have some bugs! Please raise a GitHub issue if you encounter one.

All commands assume you are in the repository root with the asd environment active.

1. Vector-graphics circle fitting (Slang + automatic boundary loss)

   python python/examples/vector_graphics_circle_fit.py \
     --image scream \
     --num-circles 100

2. Painterly rendering

   python python/examples/painterly_rendering.py \
     --image starry \
     --primitive line \
     --n 30

   Generates stylized renderings with differentiable vector graphics primitives. Results are written under results/painterly/.

Every script under python/examples/ exposes CLI flags via --help. Minimal invocation commands are listed below for completeness:

If you use this codebase in academic work, please cite:

@article{Belhe:2025:AutomaticSamplingDiscontinuities,
  author = {Yash Belhe and Ishit Mehta and Wesley Chang and Iliyan Georgiev and Micha\"{e}l Gharbi and Ravi Ramamoorthi and Tzu-Mao Li},
  title = {Automatic Sampling for Discontinuities in Differentiable Shaders},
  journal = {ACM Transactions on Graphics (Proceedings of SIGGRAPH Asia)},
  year = {2025},
  volume = {44},
  number = {6},
  doi = {10.1145/3763291}
}

Released under the MIT License. Contributions are welcome via pull requests.