💻 computer science

Grids Often Outperform Implicit Neural Representations at Compressing Dense Signals

This paper demonstrates that for compressing dense signals, simple regularized grids with interpolation often outperform Implicit Neural Representations (INRs) in both training speed and quality, while identifying specific niches like binary shape fitting where INRs remain advantageous.

Original authors: Namhoon Kim, Sara Fridovich-Keil

Published 2026-06-19

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Namhoon Kim, Sara Fridovich-Keil

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

The Big Picture: The "Mosaic" vs. The "Magic Paintbrush"

Imagine you have a giant, incredibly detailed photograph of a forest, a city, or a human lung. You want to shrink this image down to fit on a tiny USB drive without losing too much quality. You have two main tools to do this:

The Grid (The Mosaic): This is like a giant checkerboard. You divide the image into millions of tiny squares. For each square, you just write down the average color. To see the image again, you look at the square and fill it in. It's simple, predictable, and works great if the image is just a mix of colors and textures (like a forest or a lung scan).
The INR (The Magic Paintbrush): This is a "neural network." Instead of a grid, you have a smart computer program (a recipe) that says, "If you are at coordinate X, Y, Z, here is exactly what color the pixel should be." It's like a magic paintbrush that can theoretically draw any shape perfectly, no matter how zoomed in you get.

The Paper's Main Claim:
For a long time, everyone thought the "Magic Paintbrush" (INRs) was the future because it sounded smarter and more flexible. However, this paper ran a massive race between the two. They found that for dense signals (images that are full of detail everywhere, like natural photos or medical scans), the simple Grid (Mosaic) is actually faster, easier to train, and often produces a clearer picture than the Magic Paintbrush, even when they both use the same amount of memory.

The Race: How They Tested It

The researchers didn't just look at one picture. They created a "track" with different types of challenges:

The Smooth Hills (Bandlimited Signals): These are images that look like rolling hills or static noise. They don't have sharp edges; they just have smooth gradients.
The Sharp Edges (Spheres and Fractals): These are images with distinct shapes, like a ball floating in space or a fractal pattern (a shape that repeats itself infinitely, like a snowflake).
Real Life: They tested on real photos (DIV2K dataset), 3D dragon models, and human CT scans (X-rays of a chest).

They tested these tools at different "sizes" (how much memory the tool is allowed to use), ranging from very small (highly compressed) to very large.

The Results: Who Won?

1. The "Dense" Signal Winner: The Grid

When the signal was "dense" (meaning it had detail everywhere, like a noisy forest or a CT scan), the Grid won almost every time.

The Analogy: Imagine trying to describe a cloudy sky. The Grid just says, "Top left is light blue, bottom right is dark blue." It does this perfectly and instantly.
The Magic Paintbrush (INR) Struggle: The INR tried to learn a complex mathematical formula to describe the clouds. It took much longer to learn, and even with the same amount of "brain power" (parameters), it often produced blurry or wavy artifacts (weird patterns) instead of a clean image.
The Finding: For these types of signals, the simple Grid is not just "good enough"; it is actually better and faster.

2. The "Sparse" Signal Winner: The Magic Paintbrush (Sometimes)

There was one specific scenario where the Magic Paintbrush shined: Sharp, simple shapes.

The Analogy: Imagine a picture of a single black circle on a white background.
The Grid's Struggle: The Grid has to use thousands of tiny squares to approximate that circle. The edge of the circle looks "jagged" (like a staircase) unless the grid is huge.
The Magic Paintbrush's Win: The INR can learn the formula for a perfect circle. It can draw a smooth, curved edge even with very few parameters.
The Finding: If your data is mostly empty space with a few sharp, simple shapes (like a 3D model of a dragon or a shape mask), the INR can compress it better than the Grid.

The "Bandwidth" Surprise

The paper discovered a rule about how these models handle detail, which they call "bandwidth."

Think of bandwidth as the "speed limit" for how much detail a model can see.
They found that for the Grid, the speed limit goes up steadily as you give it more memory. It's a predictable, straight line.
For the INRs, the speed limit also goes up, but they hit a wall. No matter how much you increase the size of the INR, it struggles to beat the Grid on "noisy" or "dense" images. The Grid is just naturally better at handling the chaos of real-world noise.

The "Speed" Factor

Grid: Like a calculator. You press a button, and it gives you the answer instantly. It trains (learns) very fast.
INR: Like a student trying to solve a complex math problem in their head. It takes a long time to figure out the pattern, and sometimes it gets stuck or makes weird mistakes (artifacts).
The Result: The Grid was often 10 times faster to train than the slowest INR models.

Summary: When to Use Which?

The paper concludes with a simple guide for anyone trying to compress or represent signals:

Use the Grid (Mosaic) if: You are dealing with "dense" data. This includes natural photos, medical CT scans, or any image that looks like a mix of colors and textures without clear, simple shapes. It is faster, cheaper, and gives better results.
Use the INR (Magic Paintbrush) if: You are dealing with "sparse" data. This includes things like 3D object shapes, outlines, or masks where the signal is mostly empty space with sharp, clean edges.

The Bottom Line:
Don't assume the "smartest" tool (the Neural Network) is always the best. Sometimes, the simplest tool (the Grid) is the most powerful, especially when you are trying to compress the messy, detailed reality of the world.

Technical Summary: Grids Often Outperform Implicit Neural Representations at Compressing Dense Signals

Problem Statement

Implicit Neural Representations (INRs) have recently demonstrated impressive capabilities in imaging and inverse problems, offering high perceptual quality with compact memory footprints. However, their fundamental capacity, implicit biases, and scaling behaviors remain poorly understood. This lack of clarity limits the confidence with which practitioners can deploy INRs and hinders the development of future signal representations. The core problem addressed is determining when INRs offer a genuine advantage over traditional grid-based representations and hybrid approaches, particularly regarding how different methods allocate capacity across signals of varying bandwidth and dimensionality.

Methodology

The authors conducted a comprehensive empirical study comparing diverse signal representation strategies across a suite of 2D and 3D tasks. The evaluation framework included:

Representations Compared:

Pure INRs: Fourier Feature Networks (FFN), SIREN, and WIRE.
Hybrid Models: GA-Planes and Instant-NGP (multi-resolution hash encoding).
Discrete/Explicit: Gaussian Splatting (restricted to 2D in this study).
Adaptive Bandwidth: BACON.
Baseline: A simple regularized grid with interpolation (bicubic for 2D, trilinear for 3D).

Datasets and Signals:

Synthetic Signals: Generated to test varying effective bandwidths (0.1 to 0.9) and structural complexity. These included Bandlimited noise, Spheres (randomly arranged disks/balls), Sierpinski fractals, and a radial Star Target.
Real-World Signals: DIV2K high-resolution images, 2D CT scans (chest and radiotherapy planning), and 3D Stanford Dragon models (both volumetric occupancy and surface representations).

Tasks:
The models were evaluated on:

Overfitting: Fitting signals with a fixed parameter budget ranging from $1 \times 10^4$ to $3 \times 10^6$ parameters.
Inverse Problems: Computed Tomography (CT) reconstruction, denoising, and super-resolution (SR).

Metrics:
Performance was quantified using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) for 2D tasks, and PSNR and Intersection over Union (IoU) for 3D tasks. Training and inference times were also measured to assess computational efficiency.

Key Contributions

Quantified Capacity Scaling: The study establishes that for 2D bandlimited signals, most models exhibit an approximate power-law relationship between model size and the effective bandwidth they can represent.
Identification of Grid Superiority: The authors demonstrate that for many tasks involving dense signals (specifically bandlimited noise and natural images), a simple regularized grid with interpolation trains faster and achieves higher or comparable quality than any INR or hybrid model with the same number of parameters.
Definition of INR Advantage Zones: The paper identifies specific, limited scenarios where INRs and hybrid models outperform grids. These are primarily tasks involving signals with underlying lower-dimensional structures, such as shape occupancy masks, object surfaces, or signals with constant-value regions and sharp edges (e.g., the 3D Dragon surface or Sierpinski fractals).

Results

Overfitting Dense Signals:

Bandlimited Signals: For both 2D and 3D bandlimited signals, the simple Grid baseline consistently outperformed all other representations across all model sizes. INRs and hybrid models failed to reliably surpass the grid, suggesting they are not a universal solution for dense, noise-like signals.
Natural Images: On the DIV2K dataset, the Grid baseline was competitive with other methods, often matching or exceeding their performance in overfitting and super-resolution tasks.
Artifacts: Different parameterizations induced distinct artifacts. FFNs and SIRENs exhibited sinusoidal or aliasing-like artifacts; WIRE introduced texture artifacts in 3D; Instant-NGP suffered from noisy artifacts due to hash collisions at small scales; and GA-Planes showed axis-aligned artifacts. The Grid baseline primarily suffered from blurring of super-Nyquist details under strong compression.

Inverse Problems:

CT and Denoising: In computed tomography and denoising tasks involving natural 2D signals, the simple Grid with Total Variation (TV) regularization was near-optimal across all model sizes, training faster and achieving higher quality than other methods.
3D Super-Resolution: In 3D super-resolution tasks (Stanford Dragon occupancy and surface), INRs (specifically GA-Planes and WIRE) and hybrid models outperformed the Grid at the smallest model sizes. This aligns with the finding that INRs excel when the signal possesses a lower-dimensional structure (e.g., a 2D surface embedded in 3D space).

Computational Efficiency:

The Grid model provided the fastest inference and training times.
Pure INRs required the most computation time (approximately 10x that of the next slowest model in some cases).
Hybrid and discrete models fell in between, though their training times were generally independent of model size, unlike pure INRs which showed an inherent increase in computation time with model size.

Significance and Claims

The paper claims to provide practical guidance for selecting signal representations based on the specific characteristics of the data and the application.

For Dense Signals: The authors conclude that a simple grid with interpolation remains the most practical and effective choice for a wide range of applications involving dense natural signals (like CT scans and natural images). It offers simplicity, interpretability, computational efficiency, and often superior reconstruction quality.
For Structured Signals: INRs and hybrid models provide distinct advantages specifically for signals with simple underlying lower-dimensional structures, such as object edges, surfaces, or occupancy masks.
Future Direction: The work suggests that current INRs are not meeting the goal of compressing "dense" natural signals effectively, implying that future development should focus on understanding and exploiting the specific structural biases that allow INRs to succeed in sparse or structured domains, rather than treating them as a general-purpose replacement for grids.

The study emphasizes that the "infinite resolution" claim of INRs does not translate to superior performance in all regimes, particularly when compared against the theoretical guarantees of sampling theory applied to regular grids.