💻 computer science

Characterizing and Optimizing the Spatial Kernel of Multi Resolution Hash Encodings

This paper introduces a physical characterization of Multi-Resolution Hash Encodings (MHE) using Point Spread Functions to reveal how spatial resolution and anisotropy are governed by optimization dynamics and hash collisions, ultimately proposing a "Rotated MHE" (R-MHE) to mitigate these effects.

Original authors: Tianxiang Dai, Jonathan Fan

Published 2026-02-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tianxiang Dai, Jonathan Fan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to paint a hyper-realistic portrait of a person using only a set of specialized brushes.

In the world of AI, specifically in "Neural Graphics" (which helps computers create 3D models of objects and scenes), there is a very popular tool called Multi-Resolution Hash Encoding (MHE). Think of MHE as a massive box of brushes. Some brushes are tiny and sharp (for painting eyelashes), while others are huge and fluffy (for painting the background sky).

However, there has always been a problem: nobody really knew exactly how these brushes worked together. People were just "guessing" which brushes to use, leading to blurry images or weird digital "noise."

This paper, written by researchers at Stanford, acts like a microscope and a manual for these brushes. Here is the breakdown of what they discovered and how they fixed it.

1. The "Blurry Brush" Problem (The PSF)

The researchers looked at what happens when you try to paint a single, tiny, perfect dot using these brushes. In physics, this is called a Point Spread Function (PSF).

They discovered two big surprises:

The "Average" Rule: You might think that if you have one super-sharp brush, your painting will be super sharp. But they found that the "sharpness" of your final image isn't determined by your sharpest brush, but by the average of all your brushes. It’s like trying to draw a fine line with a hand that is shaking; the shaking (the coarser brushes) blurs out the precision of the sharpest pen.
The "Grid" Problem (Anisotropy): Because these digital brushes are laid out on a square grid (like graph paper), they aren't equally good in all directions. They are great at painting straight up-and-down or left-to-right, but they get "blurry" and "stretched" when you try to paint diagonally.

2. The "Ghost in the Machine" (Hash Collisions)

Because these digital brushes are stored in a limited "storage box" (a hash table), sometimes two different parts of the painting try to use the exact same brush at the same time.

Imagine you are painting a person's nose, but because of a mistake, the computer accidentally uses the same color and texture you used for a tree in the background. Suddenly, a "ghostly" speck of tree texture appears on the person's nose. This is called a collision, and it creates "speckle noise" that ruins the image.

3. The Solution: "Rotated MHE" (The Magic Spin)

To fix the "diagonal blur" problem, the researchers came up with a brilliant, simple trick called R-MHE.

If the problem is that the brushes are all aligned to a square grid, the solution is to spin the brushes.

Instead of every layer of the painting using the same square grid, they tell the computer: "For the first layer, use the grid normally. For the second layer, tilt the grid 45 degrees. For the third, tilt it again."

The Analogy: Imagine you are trying to weave a basket. If you only weave strictly up-and-down and left-to-right, the basket might be a bit stiff and uneven. But if you weave some strands diagonally and some straight, the basket becomes much more circular, sturdy, and smooth in every direction.

Why does this matter?

By understanding the "physics" of these digital brushes, the researchers did two things:

They gave us a recipe: Instead of guessing, we can now use math to pick the perfect settings to get the sharpest possible image.
They made it better for free: Their "Rotation" trick makes the 3D models look better and cleaner without requiring more memory or making the computer work harder.

In short: They turned a "guessing game" into a "science," making 3D digital worlds sharper, cleaner, and more realistic.

Technical Summary: Characterizing and Optimizing the Spatial Kernel of Multi-Resolution Hash Encodings

Published at: ICLR 2026
Authors: Tianxiang Dai & Jonathan Fan (Stanford University)

1. Problem Statement

Multi-Resolution Hash Encoding (MHE), popularized by Instant-NGP, is a cornerstone of modern neural fields (NeRFs, SDFs). While highly efficient, its performance is extremely sensitive to hyperparameters such as the number of levels ( $L$ ), growth factor ( $b$ ), and hash table capacity ( $T$ ). Currently, these parameters are selected using heuristics rather than principled engineering. The authors identify a critical gap: MHE lacks a rigorous physical understanding of its spatial behavior, leading to unexpected issues like anisotropy (direction-dependent resolution) and optimization-induced broadening (where the actual resolution is much lower than the theoretical maximum).

2. Methodology

The authors approach MHE through the lens of physical systems, treating the encoding as a system with a specific impulse response.

Point Spread Function (PSF) Analysis: They characterize the encoding by measuring its PSF—the model's response when optimized to represent an idealized point source. This is analogous to finding the Green’s function of a physical system.
Linearized Decoder Regime: To isolate the encoding's properties from the MLP's non-linearities, they analyze the system in a linearized framework (motivated by Neural Tangent Kernel theory).
Theoretical Derivation: They derive a closed-form approximation for the "collision-free" PSF, revealing that it follows a logarithmic radial decay and is shaped by a cubic B-spline kernel (the auto-correlation of the interpolation kernel).
Modeling Optimization Dynamics: They account for "spectral bias" (the tendency of optimizers like Adam to learn low frequencies first) by modeling the weights of different resolution levels as a power law, explaining why the empirical PSF is broader than the idealized one.
Collision Modeling: They formalize how finite hash capacity $T$ introduces "speckle noise" and degrades the Signal-to-Noise Ratio (SNR).

3. Key Contributions

Analytical Characterization: A mathematical derivation of the MHE PSF, proving that the idealized spatial bandwidth (FWHM) is determined by the average resolution ( $N_{avg}$ ) rather than the finest resolution ( $N_{max}$ ).
Discovery of Optimization Broadening: They prove and empirically demonstrate that gradient-based optimization induces a spatial broadening factor ( $\beta_{opt}$ ), meaning the effective resolution is significantly lower than the architecture's specification.
Rotated MHE (R-MHE): A novel, parameter-free architecture that applies distinct rotations to input coordinates at each resolution level. This breaks the axis-aligned grid dependency and restores isotropy.
Principled Hyperparameter Selection: A methodology to select the growth factor $b$ by calculating it such that the empirical FWHM matches the target spatial resolution (e.g., pixel size).

4. Results

Validation of Theory: Numerical experiments confirm that the empirical PSF matches the predicted logarithmic decay and that the FWHM scales predictably with $N_{avg}$ .
2D Image Regression: R-MHE significantly outperforms standard MHE, achieving a +0.94 dB improvement in PSNR with zero additional parameters or computational overhead.
3D NeRF & SDF:
- In NeRF, the "PSF-Guided" strategy for selecting $b$ consistently identifies the optimal performance regime, matching empirical grid searches.
- In SDFs, while performance reaches saturation at high resolutions, R-MHE maintains competitive results.
Robustness: They show that the broadening factor is remarkably stable across different MLP depths and dimensions, depending primarily on the choice of optimizer (e.g., $\beta_{emp} \approx 3.0$ for Adam).

5. Significance

This work shifts the paradigm of MHE optimization from trial-and-error heuristics to a physics-based principled approach. By providing a mathematical framework to understand resolution, anisotropy, and noise, it enables researchers to design more efficient and predictable neural field architectures. The proposed R-MHE offers a simple yet powerful way to improve reconstruction quality in memory-constrained or high-precision applications without increasing the model's footprint.