Fast Voxelization and Level of Detail for Microgeometry Rendering

Javier Fabre; Carlos Castillo; Carlos Rodriguez-Pardo; Jorge Lopez-Moreno

cs.GRApr 14, 2026

Fast Voxelization and Level of Detail for Microgeometry Rendering

Authors: Javier Fabre, Carlos Castillo, Carlos Rodriguez-Pardo, Jorge Lopez-Moreno

Organizations: Universidad Rey Juan Carlos · Politecnico di Milano · Euro-Mediterranean Center on Climate Change (CMCC) · RFF-CMCC European Institute on Economics and the Environment (EIEE)

Abstract

Many materials show anisotropic light scattering patterns due to the shape and local alignment of their underlying micro structures: surfaces with small elements such as fibers, or the ridges of a brushed metal, are very sparse and require a high spatial resolution to be properly represented as a volume. The acquisition of voxel data from such objects is a time and memory-intensive task, and most rendering approaches require an additional Level-of-Detail (LoD) data structure to aggregate the visual appearance, as observed from multiple distances, in order to reduce the number of samples computed per pixel (E.g.: MIP mapping). In this work we introduce first, an efficient parallel voxelization method designed to facilitate fast data aggregation at multiple resolution levels, and second, a novel representation based on hierarchical SGGX clustering that provides better accuracy than baseline methods. We validate our approach with a CUDA-based implementation of the voxelizer, tested both on triangle meshes and volumetric fabrics modeled with explicit fibers. Finally, we show the results generated with a path tracer based on the proposed LoD rendering model.

Figures & tables

Figure 1: Our pipeline process from sparse 3D vertices to LoD of voxels in representing 3D generated model. We generate new vertices subdividing our original primitives, each of these vertices with all information needed to be processes in the GPU. Our 3rd step gathers all valid samples already arranged in the corresponding voxel. Last, we process each generated histogram per voxel (for orientation data) and generate our whole hierarchical structure for rendering.

Parameters	Type	Bytes
2nd level index	integer	4
3rd level index	integer	4
Tangent data	u–integer	3x1
Normal data	u–integer	3x1
Material ID	u–integer	4

Table 1: Required space for our initial information per node. This data is used to compute the information that we stored in the final leaf nodes (in our examples, the voxels at the 2nd resolution level).

Panel of Figure 2 — Figure 2: Multiple comparisons for vegetation models from Hasselgen Hasselgren et al. (2021) dataset (Hibiscus and Gardenia) and grass model from free3d.com . We provide resolution data and F LIP metrics for these scenes in our Supplementary Material.

Panel of Figure 3 — Figure 3: Comparisons of our Level of Detail method against Naive downsampling for Yuksel’s hair model. We translated the original files to splines to fit our pipeline and then proceed to voxelize them to generate a Ground Truth render from the original voxelized volume. We provide resolution data and F LIP metrics for these scenes in our Supplementary Material.

Panel of Figure 4 — Figure 4: Comparisons of renders at different voxelization LoDs of the Scarf volume Jakob et al. (2010) using Ground Truth data (left) and our hierarchical model (right) for tangent data.

Panel of Figure 5 — Figure 5: Render comparison at multiple LoDs of an anisotropic steel table. Left: A three-level resolution voxelized model. Middle: LoD based on naïve average. Right: Our SGGX-H model. In the naïve approach, averages tangent orientations are collapsed into two almost perpendicular directions. Our method evaluates samples of mutliple distributions without averaging them, producing better results, even at coarse resolutions.

Panel of Figure 6 — Figure 6: Comparisons of renders at different voxelization LoDs of a helmet mesh using Ground Truth data (left), naive fit to SGGX model (center) and our hierarchical model (right) for orientation storage. We also show averaged normal per pixel. Note, however, that for both SGGX and SGGX-H we do not evaluate this averaged normal while rendering, but sample normals from the distribution for each shading evaluation.

Scene (Resolution)	Level of Detail	Method
Scene (Resolution)	(Resolution)	Naive	SGGX	Ours
Table (2048)	1 (256)	0.095	0.094	0.085
	2 (128)	0.105	0.103	0.094
	3 (64)	0.111	0.118	0.101
Helmet (512)	1 (64)	0.100	0.103	0.098
	2 (32)	0.163	0.154	0.151

Table 2: While a straightforward SGGX fit sometimes fail to fully capture aggregated our method (SGG-H) scores improved ( F LIP) metrics compared to a Naive approach for all scenes, at different Levels of Detail.

Figure 7: Comparison of SGGX-H relative improvement over Naive rendering for the Helmet model. Left column: Side-by-side renderings for selected Levels of Detail (LoDs 1, 4, 7, 11 indicated by colored borders). Right columns: Plots showing the percentage improvement of SGGX-H compared to Naive across normalized LoD for L1, LPIPS, and F LIP error metrics. The x-axis represents normalized LOD, where 0 corresponds to the lowest detail (highest LoD number) and 1 corresponds to the highest detail (lowest LoD number). Positive percentages indicate SGGX-H yields lower error than the Naive approach for the respective metric.

Scene	Parameters				Time (s)
Scene	Nodes	Fibers	Subdivisions	Resolution (1st x 2nd)	Raster	Ours	w/ Histogram
Fabric flat	8,359,496	150	60	32 x 32	81.117	3.450	8.813
				32 x 64	15.677	6.058	9.151
				32 x 128	41.996	22.339	9.396
Fabric folded	8,227,286			32 x 32	6.649	3.521	7.987
				32 x 64	14.502	6.156	8.2062

Table 3: Using our voxelizer in CUDA, we measured the times to replicate the behavior of previous approaches. Measured time include loading stage, voxelization, and exportation to OpenVDB format. We compare our voxelization times with and without histogram computation against a raster based implementation using the same GPU (NVIDIA GeForce - GTX 1080 Ti). Visualization of the fabrics voxelized for these test are shown in the Supplementary Material.

Physically Based Rendering

Physically based rendering models how light interacts with geometry and materials to produce images for synthesis, relighting, and reconstruction.

See all 41 papers

Gaussian Volumetric Representation for Efficient Shear-Warp Visualization
Jul 28, 2026Mayuri Mathur, Ojaswa SharmaScientific Visualization Medical Image Representation Learning
Indraprastha Institute of Information Technology Delhi · Indraprastha Institute of Information Technology Delhi, New Delhi, India
Bake It Till You Make It: Ultrafast Spatial Texture-Atlas Splatting
Jul 15, 2026Neel Kelkar, Simon Niedermayr, Kaloian Petkov +23DGS Acceleration Neural Rendering
Technical University of Munich, Siemens-Healthineers, Munich, Germany · Siemens-Healthineers, Erlangen, Germany
FilterGS: Traversal-Free Parallel Filtering and Adaptive Shrinking for Large-Scale LoD 3D Gaussian Splatting
Mar 25, 2026Yixian Wang, Haolin Yu, Jiadong Tang +43D Gaussian Splatting 3DGS Acceleration
Beijing Institute of Technology

Jul 28, 2026J/K move · Enter open · S save

Gaussian Volumetric Representation for Efficient Shear-Warp Visualization

Mayuri Mathur, Ojaswa Sharma

Indraprastha Institute of Information Technology Delhi · Indraprastha Institute of Information Technology Delhi, New Delhi, India

Scientific Visualization Medical Image Representation Learning

Open paper AskPDF arXiv

Medical image visualization requires volumetric rendering algorithms that preserve anatomical fidelity while maintaining high rendering speeds. To address the high computational cost of large volumetric datasets, we propose a Gaussian-based volumetric representation for efficient visualization of dense medical volumes without compromising structural and radiometric details. We optimize the proposed representation using Monte Carlo volumetric estimation, which enables training on a highly sparse subset of voxels while maintaining consistency with the dense volumetric objective. In addition, we introduce a curriculum learning strategy that progressively incorporates structured slice-based sampling during training. Sparse voxel samples provide an early global coverage of the volume, while slice samples capture spatially correlated regions that aid geometric structure and texture continuity. This combination enables the Gaussian representation to learn anatomical details of various structures and corresponding textures from sparse supervision while significantly reducing the computational cost associated with dense voxel processing. The learned representation supports slice-based rendering methods such as shear-warp volume rendering, enabling efficient visualization of multimodal medical datasets including MRI and Cryosection volumes while preserving anatomical structures. Using sparse supervision, our method achieves up to 43.86 FPS rendering with a compression ratio of 11.31:1.

Abstract

Physically Based Rendering

Explore similar work

Gaussian Volumetric Representation for Efficient Shear-Warp Visualization

Bake It Till You Make It: Ultrafast Spatial Texture-Atlas Splatting

FilterGS: Traversal-Free Parallel Filtering and Adaptive Shrinking for Large-Scale LoD 3D Gaussian Splatting