Neural Visibility Cache for Real-Time Light Sampling

2025-06-06

AI Summary

To address the efficiency bottleneck in visibility queries for multi-light direct illumination in real-time rendering, this paper proposes an online-trainable Neural Visibility Cache (NVC). The method integrates light-point visibility modeling into the Weighted Reservoir Sampling (WRS) framework and, for the first time, combines multi-resolution hash grid encoding with a fully connected MLP to achieve low-latency, high-accuracy visibility estimation on GPU. NVC supports end-to-end online training and inference, and seamlessly integrates with advanced sampling techniques such as ReSTIR. Experimental results demonstrate real-time frame rates on modern GPUs, substantial improvements in visual quality and convergence speed for multi-light illumination, and full compatibility with mainstream real-time rendering pipelines.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Multi-instance/Multi-view Learning
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Direct illumination with many lights is an inherent component of physically-based rendering, remaining challenging, especially in real-time scenarios. We propose an online-trained neural cache that stores visibility between lights and 3D positions. We feed light visibility to weighted reservoir sampling (WRS) to sample a light source. The cache is implemented as a fully-fused multilayer perceptron (MLP) with multi-resolution hash-grid encoding, enabling online training and efficient inference on modern GPUs in real-time frame rates. The cache can be seamlessly integrated into existing rendering frameworks and can be used in combination with other real-time techniques such as spatiotemporal reservoir sampling (ReSTIR).

Problem

• Real-time light sampling in physically-based rendering
• Online-trained neural cache for visibility storage
• Integration with existing real-time rendering frameworks

Innovation

• Online-trained neural cache for visibility
• Multi-resolution hash-grid encoding MLP
• Integrates with WRS and ReSTIR