Multi-Resolution Hash Grid

The grid stores learnable feature vectors at multiple spatial resolutions simultaneously. For example, a 3D coordinate is queried across 16 distinct grid levels, from coarse (e.g., 16^3) to fine (e.g., 512^3). This allows the model to capture both broad scene geometry and intricate high-frequency details, similar to a wavelet decomposition. The final feature for a point is a concatenation of the interpolated vectors from all levels.

Instead of a dense, memory-prohibitive voxel grid, each resolution level is implemented as a fixed-size hash table. A spatial hash function maps a 3D grid cell index to an entry in the table. This provides O(1) lookup time and constant memory overhead, independent of scene extent. For example, a hash table with 2^19 entries can index a far larger effective grid, enabling the encoding of unbounded scenes.

Hash collisions—where multiple spatial locations map to the same table entry—are not only allowed but are a designed feature. The model's small MLP decoder learns to disambiguate these collisions from the multi-resolution context. This acts as a form of lossy compression, where the network prioritizes allocating unique features to regions of space with complex signals, trading perfect uniqueness for extreme memory efficiency.

To query a continuous 3D location, the grid performs trilinear interpolation between the feature vectors at the 8 corners of the enclosing voxel at each resolution level. This interpolation is fully differentiable, allowing gradients to flow back to update the stored feature vectors during training. This smoothness is crucial for representing continuous scenes without blocky artifacts.

In Instant Neural Graphics Primitives, the multi-resolution hash grid replaces the traditional positional encoding used in vanilla NeRF. The concatenated features from all grid levels are fed directly into a tiny MLP (often just 1-2 layers) to predict color and density. This architectural shift is the primary reason for Instant NGP's orders-of-magnitude faster training (seconds/minutes) and real-time rendering capability.

While foundational to Instant NGP, the hash encoding principle is a general-purpose tool for accelerating coordinate-based networks. It has been successfully applied to other tasks:

• Signaled Distance Fields (SDF) for geometry.
• Neural Image Compression by encoding patches.
• Gigapixel Image Representation. Its efficiency makes it ideal for any task requiring a neural network to learn a mapping from multi-dimensional coordinates to an output signal.

Hash encoding is the memory-efficient feature indexing technique at the heart of a multi-resolution hash grid. It maps spatial coordinates to learnable feature vectors stored in a hash table.

• Key Mechanism: Uses a spatial hash function to compute an index into a fixed-size table, tolerating hash collisions gracefully via learned parameters.
• Purpose: Replaces large, dense feature grids with a compact table, enabling the representation of fine details without a cubic memory explosion.
• Example: In Instant NGP, a 3D coordinate is hashed at 16 resolution levels, retrieving 16 feature vectors that are concatenated and fed into a tiny MLP.

Instant Neural Graphics Primitives (Instant NGP) is the seminal framework that introduced the multi-resolution hash grid for real-time neural rendering.

• Architecture: Combines a multi-resolution hash grid encoder with a small, fully-connected neural network (MLP).
• Performance: Achieves training times of seconds to minutes and real-time rendering at high resolution, orders of magnitude faster than original NeRF.
• Impact: Enabled practical applications of neural radiance fields by solving the speed and quality trade-off through efficient parametric encoding.

A coordinate-based MLP is a fully-connected neural network that takes spatial (or spatio-temporal) coordinates as direct input and outputs a scene property, forming the decoder in many implicit neural representations.

• Function: Acts as a continuous function approximator, mapping (x, y, z, θ, φ) to (RGB, σ) in a standard NeRF.
• Limitation: Pure MLPs struggle with high-frequency details, a problem addressed by input encodings like positional encoding or hash grids.
• Role in Hash Grids: In Instant NGP, the MLP is intentionally kept very small (e.g., 2 layers), as the multi-resolution hash grid provides a powerful, learned input encoding.

Positional encoding is a precursor technique to hash encoding for helping neural networks learn high-frequency functions. It projects input coordinates into a higher-dimensional space using a fixed set of sinusoidal functions.

• Mathematical Form: γ(p) = (sin(2^0 π p), cos(2^0 π p), ..., sin(2^{L-1} π p), cos(2^{L-1} π p)).
• Comparison to Hash Grids: While positional encoding uses a fixed, analytical mapping, a multi-resolution hash grid uses a learned, parametric mapping. Hash grids typically offer higher quality and faster convergence for a given computational budget.
• Use Case: Remains widely used in Transformer architectures and was a key component in the original NeRF paper.

Neural Radiance Fields (NeRF) is the foundational coordinate-based neural representation that models a scene as a continuous volumetric field, mapping a 3D point and viewing direction to color and density.

• Core Contribution: Introduced the paradigm of optimizing an implicit neural representation via differentiable volume rendering.
• Relation to Hash Grids: The original NeRF used a large coordinate-based MLP with positional encoding. The multi-resolution hash grid was developed as a far more efficient drop-in replacement for this encoding stage, leading to frameworks like Instant NGP.
• Output: Rendered via the volume rendering integral.

Differentiable rendering is the framework that enables gradient-based optimization of scene parameters (like a hash grid's features) by making the image synthesis process differentiable with respect to those parameters.

• Critical Enabler: Allows losses computed on 2D rendered images to propagate backwards to update the 3D scene representation.
• Primary Technique for NeRFs: Uses the volume rendering integral and ray marching to create a differentiable pipeline.
• Application: The multi-resolution hash grid is trained end-to-end within a differentiable rendering loop, where gradients from photometric loss update the hash table's feature vectors.