Neural Importance Sampling of Many Lights

2025-05-16

AI Summary

In Monte Carlo rendering with massive light sources, conventional importance sampling suffers from low efficiency and high variance. To address this, we propose a neural-network-based method for estimating spatially varying light selection distributions. Our key contributions are: (1) the first online KL-divergence minimization training scheme, enabling end-to-end optimization of the distribution estimator; (2) tight coupling with a hierarchical light structure to support cluster-level distribution prediction and efficient querying; and (3) a residual learning architecture that accelerates convergence and maintains compatibility with existing samplers. Experiments demonstrate that our method significantly reduces rendering variance in complex scenes with hundreds to thousands of lights, consistently outperforming state-of-the-art approaches in image quality while enabling real-time, high-throughput sampling.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Machine Learning: Calibration & Uncertainty Quantification
• Computer Vision: Diffusion Models for Vision

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

We propose a neural approach for estimating spatially varying light selection distributions to improve importance sampling in Monte Carlo rendering, particularly for complex scenes with many light sources. Our method uses a neural network to predict the light selection distribution at each shading point based on local information, trained by minimizing the KL-divergence between the learned and target distributions in an online manner. To efficiently manage hundreds or thousands of lights, we integrate our neural approach with light hierarchy techniques, where the network predicts cluster-level distributions and existing methods sample lights within clusters. Additionally, we introduce a residual learning strategy that leverages initial distributions from existing techniques, accelerating convergence during training. Our method achieves superior performance across diverse and challenging scenes.

Problem

• Improving light selection for Monte Carlo rendering in complex scenes
• Predicting light distributions using neural networks and local information
• Integrating neural approach with light hierarchy for efficient light management

Innovation

• Neural network predicts light selection distributions
• Combines neural approach with light hierarchy techniques
• Residual learning accelerates training convergence