Scale-Aware Isotropy for Multi-Resolution Hash Grids

Many fast neural fields store features on a stack of axis-aligned grids at increasing resolution and hash the finest ones into a fixed-size table. The best-known example is the multi-resolution hash encoding of Instant-NGP. Such a grid blurs more along diagonals than along the coordinate axes: its effective blur kernel, or point spread function, is diamond-shaped rather than round. A recent fix, Rotated MHE, gives each grid level its own fixed rotation so that the diamonds average into a disc Characterizing and Optimizing the Spatial Kernel of Multi Resolution Hash Encodings.

That averaging only works when many levels contribute. Anti-aliased renderers do the opposite at coarse scales. When a pixel covers a large patch of the scene, because the camera is far away or the image is downsampled, they deliberately suppress the fine levels. Only a few coarse levels remain, and a few arbitrarily chosen rotations do not average into a disc. We predict that grid-induced directional bias returns exactly at coarse viewing scales, and that this bias is invisible in the single-scale benchmarks where Rotated MHE was evaluated.

The proposal has two parts. The first derives, rather than tunes, how much weight each level should receive for a given pixel footprint. The second chooses the per-level rotations so that every contiguous band of levels a footprint can activate is itself close to isotropic. We will then test whether this closes a measurable scale-dependent anisotropy gap on multiscale benchmarks.

Core Claim

At scale $\sigma$, meaning the pixel footprint expressed in scene units, an encoding with per-level weights $w_l(\sigma)$ has the effective kernel

$$K_\sigma(x) = \sum_{l=1}^{L} w_l(\sigma)\, k_l(R_l x),$$

where $k_l$ is the diamond-shaped kernel of level $l$ and $R_l$ is that level's rotation. Rotated MHE chooses the $R_l$ to make $K_\sigma$ isotropic when all weights are equal. Our claim is that it should instead be isotropic for every $\sigma$ the renderer will encounter.

In 2D the diamond has four-fold symmetry, so only angles modulo $90^\circ$ matter. Summing $m$ copies rotated by evenly spaced offsets of $90^\circ/m$ cancels every angular harmonic of the kernel below order $4m$. A single fixed sequence cannot be evenly spaced for every window of consecutive levels. A golden-ratio angle sequence taken modulo $90^\circ$ comes close: by the three-distance theorem, any run of consecutive terms splits the circle into gaps of at most three distinct lengths. This makes every window of levels near-uniform in angle. In 3D the analogous choice is a low-dispersion sequence over rotations modulo the cube's symmetry group.

The weights $w_l(\sigma)$ come from the same analysis. That paper shows the kernel width is set by the average level resolution rather than the finest one, and that the optimizer broadens it by roughly $\beta_{opt}\approx 2.5$. We choose $w_l(\sigma)$ so that the predicted width of $K_\sigma$, including this broadening, matches the footprint. This makes the resolution-dependent parameters a computed quantity instead of a per-dataset search.

Relation to Prior Work

The Rotated MHE paper Characterizing and Optimizing the Spatial Kernel of Multi Resolution Hash Encodings supplies the kernel analysis we reuse and motivates the gap. Its rotations are static and indexed by level, and its 3D gains were small, reaching 35.48 dB PSNR on Synthetic NeRF, which the authors attribute to averaging across views. We keep its zero-overhead rotation mechanism but tie both the rotations and the level weights to viewing scale.

Three anti-aliasing methods adapt to scale but ignore orientation. Mip-NeRF Mip-NeRF integrates its positional encoding over a Gaussian cone footprint. Mip-VoG Multiscale Representation for Real-Time Anti-Aliasing Neural Rendering interpolates across mipmapped axis-aligned voxel grids. LoD-NeuS Anti-Aliased Neural Implicit Surfaces with Encoding Level of Detail convolves tri-plane features with Gaussians sized to the viewing cone. All three decide how much to blur; none asks whether the blur they get from a grid is round. We supply that missing orientation control for hash grids.

Classical theory constrains the design. Continuous limits of discrete perimeters proves that fixed-stencil lattice energies converge to faceted, direction-dependent limits even as grid spacing shrinks, unless the stencil changes with resolution. Its analysis concerns binary lattice energies, not neural encodings, so we use it only as the analogue that rules out "just use finer grids." It tells us that the directional correction must itself depend on scale, which is exactly what $R_l$ paired with $w_l(\sigma)$ provides.

Continuous-mesh approaches such as A unifying resolution-independent formulation for early vision. remove grid bias entirely. They cost minutes per image and handle stochastic texture poorly, so we treat them as the isotropy ideal rather than as a practical competitor.

Research Program

We first build a measurement tool: a scale-resolved anisotropy score $A(\sigma)$. It is defined as the ratio of the largest to smallest directional half-maximum width of $K_\sigma$, minus one. We measure it two ways. Analytically, we evaluate the closed-form kernel. Empirically, we fit a 2D field to a point impulse and read off the width after optimization, which captures optimizer broadening.

We then compute $A(\sigma)$ for plain MHE, for Rotated MHE with its random or arbitrary rotations, and for golden-sequence rotations. Each is paired with footprint-derived weights. This analysis is cheap and settles quickly whether the predicted coarse-scale re-emergence of anisotropy is real.

Next we integrate the weighting and rotations into an Instant-NGP-style renderer that computes cone footprints the way Mip-NeRF does. Level weights are derived from the kernel-width prediction. We also run a learned-weight variant, in which a small function of $\sigma$ is trained per level, to test whether the analytic rule loses anything.

Hash collisions distort the kernel in ways the theory ignores. To keep them from confounding the orientation question, every analysis is run twice: once in a collision-free regime with a large table, and once at the standard table size.

Baselines

The naive reference is plain MHE with the same footprint-derived level weights and no rotations. It shows how much anisotropy an anti-aliased hash grid carries by default.

The strongest direct competitor is Rotated MHE Characterizing and Optimizing the Spatial Kernel of Multi Resolution Hash Encodings with its published rotation choice, combined with our footprint weighting. This is the key ablation. It differs from the full method only in how rotations are chosen, so it isolates whether window-wise isotropy matters.

We also run Rotated MHE without footprint weighting, which reproduces the published setting. The file reports 35.48 dB on single-scale Synthetic NeRF for this configuration.

As the strongest existing multiscale system, we report Mip-VoG. The file gives 30.33 dB average PSNR on Multiscale Blender for it Multiscale Representation for Real-Time Anti-Aliasing Neural Rendering. For Mip-NeRF the file gives only a 60% error reduction relative to NeRF on Multiscale Blender, and no absolute PSNR. We therefore retrain it rather than quote a number.

Two further baselines address alternatives a reviewer would raise. One applies Gaussian prefiltering of features in the style of LoD-NeuS. The other is 4× supersampled plain MHE.

All hash-based variants share the table size, level count, growth factor, iteration budget, and optimizer. The only tuning allowed is a fixed sweep over learning rate applied identically to every variant.

Beating Rotated MHE means two things together. First, lower $A(\sigma)$ at the two coarsest scales. Second, a statistically reliable PSNR gain at the 4× and 8× downsampled scales of Multiscale Blender, measured across the eight scenes with paired tests, with no loss at 1×. Beating Mip-VoG's 30.33 dB is a secondary target, since that system uses a different representation.

Why not just…?

Why not just combine Rotated MHE's random rotations with footprint weighting? This is the real threat. If the contiguous windows of levels that footprints actually activate are wide, typically six or more levels, random rotations may already cancel the diamond well enough. Our method would then add nothing. The analytic $A(\sigma)$ curves settle this before any 3D training. If random and golden rotations give indistinguishable coarse-scale anisotropy, we will drop the rotation-sequence claim. We would then report the footprint-derived weights and the scale-resolved measurement as the contribution.

Why not just prefilter features with an isotropic Gaussian, as LoD-NeuS does? Convolving a diamond with a round Gaussian of similar width makes it rounder, but it does not make it round. Removing the bias requires a filter wider than the footprint, which throws away resolution. The filtering also costs extra queries or feature convolutions, whereas rotations are free. The prefiltering baseline, plotted as sharpness against $A(\sigma)$ at matched compute, tests this directly.

Why not just use finer grids or supersample? The lattice-limit result Continuous limits of discrete perimeters shows that refinement with a fixed stencil keeps its directional bias. The hash-kernel analysis shows that kernel width follows average rather than finest resolution. Supersampling reduces aliasing, but each sample still goes through the same diamond kernel. Mip-NeRF also reports that brute-force supersampling is much slower than cone integration at matched accuracy, 22× slower against 128× supersampling. Our 4× supersampling baseline checks whether it reduces $A(\sigma)$ at all.

Why not rely on averaging across views? Rotated MHE's own small 3D gains suggest that averaging across camera angles hides the bias. It does not remove it, and at coarse scales few rays cover a structure, so there is little to average. The rotated-scene test below measures what that averaging leaves behind.

Proposed Experiments

The decisive 2D check is gigapixel image regression, where there is no averaging across views and Rotated MHE reported a +0.94 dB gain over MHE. We fit each image, then evaluate reconstruction when queried at footprints of 1×, 2×, 4×, and 8× against correspondingly low-pass-filtered ground truth. We also fit the same image rotated by 45° and by 22.5°. An isotropic encoding should show near-zero PSNR difference between rotated and unrotated fits at every scale. We report that difference as a function of scale for each variant.

In 3D we use Multiscale Blender at its four scales, reporting per-scale PSNR, SSIM, and LPIPS. We add a rotated-scene protocol: each Blender object is rotated 45° about its vertical axis before training and scored against its own renders. The rotation-sensitivity gap then measures what averaging across views leaves behind.

Single-scale Synthetic NeRF serves as a no-regression check against Rotated MHE's reported 35.48 dB.

The hypothesis fails if either of two things happens. Golden-sequence rotations may fail to lower coarse-scale $A(\sigma)$ relative to random ones. Or lower $A(\sigma)$ may not translate into smaller rotation-sensitivity gaps or better 4× and 8× PSNR. Either outcome would be informative, because it would locate grid-induced anisotropy as a real but perceptually negligible property of anti-aliased hash fields.