About sxyu/svox2

The reference implementation of Plenoxels: Radiance Fields without Neural Networks (Yu et al., 2021), distributed as the svox2 library. See the paper page for context.

Plenoxels replaces the NeRF MLP with a sparse voxel grid of spherical-harmonics coefficients optimized directly by gradient descent, achieving NeRF-quality novel view synthesis at a fraction of the training cost. The repo provides CUDA kernels for trilinear interpolation, total variation regularization, and the unbounded-scene voxel layout. Mirrored here under BSD-2-Clause — README only.