Benchmarking NeuralPVS with fVDB for higher performance and sustainability
Date
2026
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis adapts NeuralPVS, the first deep-learning approach for efficient from-region visibility computation in large scenes, from its original spconv backend to the modern, actively maintained NVIDIA fVDB backend. The adapted fVDB backend preserves the core NeuralPVS pipeline while reformulating its sparse convolutional operations into the fVDB framework. This implemented backend is then benchmarked against the original spconv backend across multiple scenes, view-cell radii, and interleaving grid sizes.
Across the majority of tested parameter configurations, the adapted fVDB backend either matches or outperforms the existing spconv backend in prediction and rendering quality. However, the fVDB backend faces challenges at the largest view-cell radius and highest interleaving grid size, where false-negative rates increase. In addition, fVDB requires higher inference time and GPU memory usage than spconv, largely due to its more explicit handling of sparse topology, feature storage, and grid transformations.
Furthermore, this thesis connects the adapted model to practical rendering workflows. A live prototype is created in the Unity pipeline for real-time rendering, and the existing VISUS Unity pipeline is converted into the NVIDIA Falcor framework. This provides a secondary, more research-oriented rendering pipeline for future NeuralPVS experimentation.