Skip to main content
Back

SGGX-H Makes Brushed Metal and Silk Look Real at Any Distance

A CUDA voxelizer and hierarchical SGGX clustering approach preserves anisotropic material appearance across Level-of-Detail hierarchies for fabrics, hair, and brushed metals.

Paper, Full Access
  • New GPU voxelizer plus SGGX-H hierarchical clustering for microgeometry rendering.
  • Solves anisotropic Level-of-Detail without destroying fiber orientation data.
  • Handles fabrics, hair, brushed metals, and vegetation through a unified pipeline.
  • CUDA implementation processes 160M+ node fabrics where raster methods run out of memory.
  • Beats naive averaging and vanilla SGGX on FLIP, LPIPS, and L1 across all LoD levels.
  • Source code and scenes to be released upon publication per the authors.

Rendering the fuzzy sheen of a silk scarf, the anisotropic streaks on brushed aluminum, or the tangle of individual strands on a character model is one of the harder problems in computer graphics. The geometry is sparse, the detail is microscopic, and every attempt to shrink the representation for distant views tends to smear away the very directional cues that make the material look like itself. A new paper from researchers at Universidad Rey Juan Carlos and Politecnico di Milano attacks both halves of that problem with a parallel voxelizer and a hierarchical clustering scheme built on the SGGX microflake distribution.

Materials with fine-scale structures like fabrics, hair, and brushed metals occupy a tiny fraction of their bounding volume yet demand high resolution to capture their detail. Storing them as dense voxel grids wastes memory, and the standard trick of averaging neighboring voxels for coarser Level-of-Detail (LoD) tiers destroys the orientation information that governs how light scatters off tiny fibers or grooves.

Why averaging breaks brushed metal

MIP mapping, the classic pyramid-of-resolutions technique, works fine for colors and roughness but fails for directional data like normals and tangents. Prior fiber-focused approaches have used SGGX (Symmetric GGX) microflake distributions to downsample volumetric data, but they inseparably mix the original directional data with scattering information at each voxel, which locks you into a specific shading model and destroys editability.