Level-of-detail decomposition of a Gabor Field: frequency- and orientation-selectivity lets us mask particle collections at render time for continuous LOD, faster rendering, and a controllable bias–variance trade-off when sample cost matters. OpenVDB volumetric Bunny.

Gaussian-based volumetric representations make physically based rendering practical at a fraction of the memory and runtime cost of dense voxel grids, but hierarchical level-of-detail (LOD) remains inneficient: prefiltering or mipmap-style Gaussians tends to cost extra memory, requires re-fitting per level, and can produce harsh transitions between LODs. We introduce Gabor Fields: mixtures of primitive anisotropic Gabor kernels (Gaussian envelopes with harmonic modulation). Limiting frequency content corresponds to pruning oriented primitives, without growing the representation. Combined with analytic line integrals and a hierarchical fitting procedure, this yields natural LOD, stochastic masking along rays for cheaper single- and multiple-scattering traversal, and allows for easy integration within volumetric path tracing frameworks.

Gabor kernels: localized in space and in frequency

Unlike Gaussians, each Gabor primitive is selective in both frequency and orientation: its spectrum is concentrated around a modulation frequency, which makes band-limiting the volume a matter of dropping kernels rather than convolving into ever-larger Gaussians.

LOD: Gaussian vs. Gabor trade-offs

Classic voxel LOD removes detail by downsampling or pruning hierarchy levels. Gaussian mixtures lack a similarly cheap spectral knob: low-pass filtering inflates supports or forces re-optimization. Gabor Fields align LOD with frequency content so coarse levels stay consistent and memory does not multiply with each mip.

Volumetric Bunny: static comparison across band-limited LOD levels (left), and camera motion (right) to illustrate temporal stability of coarse levels.

Orientation selectivity and ray-space masking

Kernels that barely contribute along a given direction can be culled or sampled stochastically, which cuts traversal work in dense volumes while remaining unbiased under our sampling strategies (see paper for derivations and variance discussion).

Selectivity of Gabor contributions in the volume.

Orientation-aware threshold culling on the Bunny.

LOD results

Far-field LOD: high-sample reference (left) vs. aggressive low-sample rendering with structured residual (right), illustrating performance-oriented trade-offs.

Accelerating multiple-scattering with our frequency-based stochastic-analytical estimators



Results





Motion blur





Acknowledgements



Affiliation