Improved Direct Voxel Grid Optimization for Radiance Fields Reconstruction

TL;DR

DVGOv2 delivers a faster, simpler direct voxel grid reconstruction for neural radiance fields by fusing CUDA-accelerated operations and re-implementing core components. It extends DVGO to forward-facing and unbounded inward-facing capture and replaces the O(N^2) distortion term with an O(N) formulation, boosting both quality and training speed. The approach relies on efficient regularization (TV) and a CUDA-accelerated rendering pipeline, with thorough ablations showing consistent speedups (~2–3x) and competitive results across bounded, forward-facing, and unbounded datasets. Overall, DVGOv2 demonstrates that dense voxel grids, when CUDA-optimized and paired with efficient regularizers, can rival faster, more complex data structures while maintaining simplicity and accessibility.

Abstract

In this technical report, we improve the DVGO framework (called DVGOv2), which is based on Pytorch and uses the simplest dense grid representation. First, we re-implement part of the Pytorch operations with cuda, achieving 2-3x speedup. The cuda extension is automatically compiled just in time. Second, we extend DVGO to support Forward-facing and Unbounded Inward-facing capturing. Third, we improve the space time complexity of the distortion loss proposed by mip-NeRF 360 from O(N^2) to O(N). The distortion loss improves our quality and training speed. Our efficient implementation could allow more future works to benefit from the loss.