06 May 2026 · 2 min · warp c64a6c3, in 1.13.0

Anisotropic voxels, or why medical volumes hate cubes

CT stacks are 6:1 anisotropic. Resampling to cubes costs 6x memory and invents data. Now in Warp 1.13.0.

A CT or MRI stack is not a stack of cubes. In-plane resolution might be 0.5 mm while slices sit 3 mm apart: a 6:1 aspect ratio baked into the acquisition itself. The scanner measured finely in two axes and coarsely in the third, and the data is honest about it.

Voxel grids that only support one scalar voxel_size are not. They force a choice between two bad resamplings:

Resample to the finest axis and pay 6x the memory, filling the slice gaps with interpolated values the scanner never measured. For a signed distance field used in collision, interpolated detail becomes fabricated geometry.
Resample to the coarsest axis and throw away the in-plane resolution you paid for.

fig 1 · the same 64 mm volume: 2.10M cubic voxels with invented slices, or 0.35M native anisotropic voxels · rendered with manim · scene source

64 mm cube at 0.5 mm isotropic: 2.10M voxels · at (0.5, 0.5, 3.0): 0.35M, no interpolation

The fix was an interface, not an engine

Warp’s volume machinery could already represent this: allocate_by_tiles() and allocate_by_voxels() support anisotropic transforms through the internal transform buffers. But the public entry points, Volume.load_from_numpy() and Volume.allocate(), assumed a scalar voxel_size and crashed on a tuple. The capability existed one layer down; the interface never exposed it.

# before: scalar only
vol = wp.Volume.load_from_numpy(sdf, voxel_size=0.5)

# after: per-axis spacing, matching the acquisition
vol = wp.Volume.load_from_numpy(sdf, voxel_size=(0.5, 0.5, 3.0))
assert diagonal(vol.grid_transform) == (0.5, 0.5, 3.0)

The patch threads the 3-vector through to the existing machinery and adds tests that verify the grid transform diagonal matches the requested spacing. It landed as commit c64a6c3 and shipped in the Warp 1.13.0 release, closing GH-1193. The GitHub PR itself shows closed, which is how NVIDIA lands external patches through its internal pipeline; the commit on main is the artifact.

The subtle part: gradients

Anisotropy does not stop at storage. A finite-difference gradient on a non-uniform grid must scale per axis:

∂f/∂z ≈ (f[k+1] − f[k−1]) / (2 · h_z), not / (2 · h)

Skip that and your SDF normals are wrong by the aspect ratio, which for a 6:1 medical stack means collision responses pointing 80 degrees off the surface. If you adopt anisotropic voxels, audit every gradient consumer the same day.

Where else this shows up

Anything with a preferred axis: depth-camera volumes where range uncertainty grows with distance while lateral resolution does not, thin-slab regions of interest, terrain where vertical detail is cheap to waste. Cubes are a convention, not a law. The data usually knows better.