Automatic Sampling for Discontinuities in Differentiable Shaders

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.

Getting Started

Prerequisites

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

Environment setup

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.

Shader compilation cache

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.

Running Examples

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

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

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

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/.

Full example command list

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

Script	Sample command	Notes
`binary_fit.py`	`python python/examples/binary_fit.py --image data/diamond_input.png`	Binary grid fitting demo.
`cel_shading.py`	`python python/examples/cel_shading.py --num-iter 400`	Optimizes cel-shading shader parameters.
`csg_opt.py`	`python python/examples/csg_opt.py`	CSG gradient optimization with boundary loss.
`ellipsoid_rasterizer_opt.py`	`python python/examples/ellipsoid_rasterizer_opt.py --scene bunny --num-iter 200`	Multi-view ellipsoid fitting.
`gaussian_mixture_fit.py`	`python python/examples/gaussian_mixture_fit.py --image scream --train-resolution 256 --save-resolution 512`	Pure PyTorch Gaussian blobs.
`grad_bezier.py`	`python python/examples/grad_bezier.py`	Gradient check for quadratic Bezier shader.
`grad_circle_single.py`	`python python/examples/grad_circle_single.py`	Gradient check for single circle shader.
`grad_circles.py`	`python python/examples/grad_circles.py`	Multi-circle gradient test.
`grad_half_plane.py`	`python python/examples/grad_half_plane.py`	Half-plane shader gradient verification.
`grad_triangle.py`	`python python/examples/grad_triangle.py`	Triangle shader gradient verification.
`grad_voronoi.py`	`python python/examples/grad_voronoi.py`	Voronoi shader gradient check.
`implicit_raymarching.py`	`python python/examples/implicit_raymarching.py --num-iter 200`	Inverse implicit ray-marching setup.
`inverse_swept_surface_2d.py`	`python python/examples/inverse_swept_surface_2d.py --num-iter 400`	2D swept surface inversion.
`painterly_rendering.py`	`python python/examples/painterly_rendering.py --image starry --primitive line --n 30`	Painterly vector graphics optimization.
`swept_surface_2d.py`	`python python/examples/swept_surface_2d.py --num-iter 200`	Forward swept-surface rendering.
`triangle_rasterizer_opt.py`	`python python/examples/triangle_rasterizer_opt.py --num-iter 200`	Triangle rasterizer optimization.
`trilinear_threshold_opt.py`	`python python/examples/trilinear_threshold_opt.py --num-iter 400`	3D threshold field fitting.
`vector_graphics_circle_fit.py`	`python python/examples/vector_graphics_circle_fit.py --image scream --num-circles 100`	Circle-only vector graphics fitting.
`voronoi_opt.py`	`python python/examples/voronoi_opt.py --image scream --grid-size 30`	Voronoi mosaic optimization.

Repository Layout

Path	Description
`python/examples/`	Standalone optimization demos (Slang and PyTorch).
`python/utils/`	Boundary sampling utilities, camera helpers, IO, and Slang runtime bindings.
`slang/`	Differentiable shader kernels. Generated shaders share the same directory with a `__gen__` prefix.
`compiler/`	Tools for compiling Slang to CUDA kernels through SlangTorch.
`results/`	Default output directory for examples.

Citation

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}
}

License

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