code for SIGGRAPH 2026 paper "Neural Quadrature Rule and Autoregressive Adaptive Sampling".

project page: https://suikasibyl.github.io/nqr

Download the shared assets from:

• https://drive.google.com/file/d/1gydL15reAH7DDNNiieg5VnS8oe0HqGDT/view?usp=sharing

After downloading, place these directories under the repo root:

• scenes/
• ckpt/

So the layout should look like:

nqr/
  ckpt/
  scenes/
  example-1d/
  example-gwn/
  example-transmittance/
  example-wos/
  example-udf/
  example-di/

You can create the Python environment with:

bash /home/haolin/Projects/nqr/scripts/create_nqr_env.sh

For the direct-illumination example, you also need SIByL Engine 0.0.5:

• https://github.com/SuikaSibyl/sibylengine

This is required for example-di/, which depends on the SIByL renderer/runtime.

• example-1d/: 1D toy integration benchmark with INet, SNet, and baseline comparisons.
• example-gwn/: generalized winding-number integration example with standalone and joint training.
• example-transmittance/: volumetric transmittance estimation with learned quadrature and adaptive sampling.
• example-wos/: walk-on-spheres style PDE example, including the non-linear p-Laplacian variants.
• example-udf/: unsigned-distance-field rendering/integration example with INet and ISNet.
• example-di/: direct illumination with SIByL-based rendering, neural quadrature, and autoregressive adaptive sampling.

Each example directory contains its own README with the local training and inference entrypoints.