suikasibyl/nqr

This repository is the code release for a SIGGRAPH 2026 research paper called Neural Quadrature Rule and Autoregressive Adaptive Sampling, written mostly in Python. Quadrature is a mathematical technique for estimating the value of an integral, and it comes up in computer graphics when simulating how light moves through a scene. The paper explores training a neural network to choose better sample points for that estimation, with the goal of making certain rendering and simulation calculations more accurate or efficient. Before running anything, a user needs to download a set of shared assets, including sample scenes and pretrained model checkpoints, from a linked Google Drive folder, then place them into specific folders inside the project so the file layout matches what the code expects. A provided shell script sets up the Python environment. One of the examples also depends on a separate renderer called SIByL Engine, version 0.0.5, which is a different project from the same author. The code is organized into six example folders, each testing the neural quadrature idea in a different setting. One is a simple one dimensional integration benchmark comparing two neural network variants against baseline methods. Another computes generalized winding numbers, a way of describing whether a point sits inside or outside a shape. A third estimates how light passes through volumes such as fog or smoke. A fourth uses a walk on spheres method, an approach for solving certain equations used in physics simulations, including harder nonlinear variants. A fifth works with unsigned distance fields, a way of representing three dimensional shapes. The last combines the neural quadrature method with the SIByL renderer for a full direct illumination lighting example. Each example folder has its own instructions for training and running its model, so someone trying this out would pick whichever example matches their interest. This is a research code release tied to a specific academic paper rather than a general purpose tool, and it is aimed at people already working in computer graphics, rendering, or neural simulation research.