Hybrid Gaussian-Hash-Grid Radiance Models

• Hybrid Gaussian-Hash-Grid Radiance Representation is a scene modeling approach that combines explicit Gaussian primitives with multiresolution hash grid features to achieve efficient and accurate rendering.
• It distinguishes methods where hash grids directly predict radiance attributes from those using grids as auxiliary context for compression, entropy coding, or editing support.
• Successful implementations balance an explicit residual pathway on Gaussians with neural hash grid components to enhance view-dependent effects while reducing storage and computational demands.

Searching arXiv for the cited works to ground the article in current literature. Hybrid Gaussian-Hash-Grid Radiance Representation denotes a class of scene models that combine explicit Gaussian primitives with multiresolution hash-grid features, but the literature uses that combination in two materially different ways. In one line of work, the hash grid directly participates in render-time radiance evaluation by predicting, modulating, or decoding Gaussian attributes such as color, opacity, scale, rotation, or residual appearance. In the other, the hash grid is auxiliary: it provides context for compression, entropy modeling, or editing support, while the decoded scene remains an ordinary Gaussian or anchor-based representation. This distinction is central to the topic, because papers such as HyRF are direct hybrid explicit–neural radiance representations, whereas HAC uses a binary hash grid as a conditional prior for entropy coding and removes it after decoding (Wang et al., 21 Sep 2025, Chen et al., 2024).

1. Scope and classification

The literature around Gaussian–hash-grid hybrids is best read as a spectrum rather than a single architecture family. Some methods replace the most memory-intensive parts of 3D Gaussian Splatting with shared hash-grid-conditioned neural prediction; some use Gaussians as editable or geometry-bearing carriers that condition a neural field; some separate low-frequency appearance on primitives from high-frequency residuals in a hash grid; and some use a hash grid only to compress an otherwise standard Gaussian model.

Method	Scene representation	Role of hash grid
Compact 3D Gaussian Representation (Lee et al., 2023)	Explicit Gaussians + neural color field	View-dependent color via Instant-NGP-style hash grid
HyRF (Wang et al., 21 Sep 2025)	Compact explicit Gaussians + decoupled neural fields	Predicts geometry residuals, view-dependent color, and background
GENIE (Zieliński et al., 4 Aug 2025)	Gaussian-conditioned NeRF	Supplies Gaussian embeddings sampled from a hash grid
Hybrid Latents (Kelkar et al., 16 Apr 2026)	2D surfels + per-surfel latents + neural field	High-frequency residual texture at ray–surfel intersections
HAC (Chen et al., 2024)	Anchor-based 3DGS compression	Conditional prior for entropy coding, removed after decoding

A recurring misconception is that every paper containing both Gaussians and a hash grid is a render-time hybrid radiance field. HAC explicitly rejects that reading: it is built on Scaffold-GS, learns a binary hash grid as structured context for entropy modeling, and restores a decoded anchor-and-Gaussian model for rendering, with the hash grid removed from the inference loop once attributes are decoded (Chen et al., 2024). By contrast, HyRF, GENIE, and the compact 3D Gaussian formulation use the hash grid during actual radiance prediction or Gaussian attribute generation (Wang et al., 21 Sep 2025, Zieliński et al., 4 Aug 2025, Lee et al., 2023).

2. Primitive substrate and common rendering model

Most direct Gaussian–hash-grid hybrids inherit the 3DGS rendering substrate. HAC restates the Gaussian primitive as

$G(\mathbf{x})=\exp\!\left(-\frac{1}{2}(\mathbf{x}-\mathbf{\mu})^\top \mathbf{\Sigma}^{-1}(\mathbf{x}-\mathbf{\mu})\right),$

with covariance parameterized by scale and rotation as

$\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$

and rendering by alpha-composited splatting,

$\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$

Those equations remain the geometric and compositing backbone even when the appearance head is replaced by a hash-grid-conditioned neural module (Chen et al., 2024).

A first important variant is the anchor formulation inherited by HAC from Scaffold-GS. Instead of storing every Gaussian independently, the scene is organized by anchors with location

$\mathbf{x}^a\in\mathbb{R}^3$

and anchor attributes

$\mathcal{A}=\{\mathbf{f}^a\in\mathbb{R}^{D^a},\mathbf{l}\in\mathbb{R}^6,\mathbf{o}\in\mathbb{R}^{3K}\},$

where $\mathbf{f}^a$ is an anchor feature, $\mathbf{l}$ is scaling, and $\mathbf{o}$ contains $K$ 3D offsets. In that representation, Gaussian centers are anchor positions plus offsets, and opacity, RGB, scale, and quaternion are predicted by MLPs from anchor attributes rather than stored independently (Chen et al., 2024).

A second variant replaces 3DGS volumetric support with surfel-style 2D Gaussians. Hybrid Latents parameterizes each surfel by position $\mu_i$, rotation $\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$0, two-dimensional scale $\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$1, and opacity $\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$2, then augments those explicit 2D primitives with a per-surfel latent and a hash-grid feature queried at the exact ray–surfel intersection (Kelkar et al., 16 Apr 2026). This preserves Gaussian-like rasterization while shifting appearance prediction into a hybrid explicit–neural latent space.

3. Render-time hybridization inside Gaussian models

The most direct form of hybrid Gaussian–hash-grid radiance representation replaces high-dimensional per-Gaussian appearance storage with a shared hash-grid neural field. Compact 3D Gaussian Representation does this for view-dependent color. The method keeps explicit Gaussian position, opacity, scale, and rotation, but replaces the $\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$3 spherical-harmonic parameters out of the $\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$4 parameters of a degree-3 3DGS Gaussian with a hash-grid-conditioned color field,

$\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$5

The paper explicitly adopts a hash-based grid representation in the style of Instant-NGP, with 16 levels, 2-channel features, resolutions from 16 to 4096, and a 2-layer 64-channel MLP. Its own ablations matter as much as its final numbers: color-only replacement with I-NGP works well, whereas using I-NGP for opacity or scale fails badly, and using it for rotation is also much worse than the proposed decomposition. The representation is therefore genuinely hybrid, but only for the radiance head; geometry remains explicit, and not all Gaussian attributes are equally amenable to grid prediction (Lee et al., 2023).

HyRF pushes the hybridization further by splitting Gaussian attributes by role and frequency. Each explicit Gaussian stores only

$\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$6

so only 8 scalar parameters per Gaussian, while decoupled geometry and radiance neural fields with Instant-NGP multi-resolution hash encoding predict the remaining properties. After contraction and encoding, geometry and color are produced by separate branches, and the explicit and neural parts are aggregated as

$\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$7

Rotation is purely neural; position is always explicit; color, opacity, and scale are residual combinations of explicit and neural terms. The paper reports that a single neural field predicting everything underperforms because geometry and appearance are weakly correlated, which is why HyRF uses separate geometry and appearance pathways, each built on a hash encoding with tiny-cuda-nn fully fused MLPs (Wang et al., 21 Sep 2025).

Hybrid Latents introduces a more explicit frequency split. Each surfel stores a learnable latent

$\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$8

while a 3D hash grid is queried at the exact 3D ray–surfel intersection point,

$\mathbf{\Sigma}=\mathbf{R}\mathbf{S}\mathbf{S}^\top\mathbf{R}^\top,$9

The pixel feature is then alpha-composited in latent space,

$\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$0

and decoded with view-direction SH,

$\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$1

The intended bias is explicit: per-surfel latents absorb low-frequency, geometry-aligned “base coat,” while the hash grid carries high-frequency residual texture. The method removes NeST’s coarse hash levels and uses a single-level hash grid, precisely to keep coarse content on the primitives and reduce geometry–appearance entanglement (Kelkar et al., 16 Apr 2026).

4. Gaussian-conditioned neural fields and editable hybrids

GENIE occupies a different but still central position in this design space. It does not splat Gaussians as final appearance elements. Instead, it uses Gaussians as editable local carriers of latent information that condition a NeRF. The model is written as

$\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$2

Gaussian color coefficients are replaced by trainable feature embeddings derived from a multi-resolution hash grid evaluated at Gaussian means, and a query point aggregates nearby Gaussian features through covariance-aware weights:

$\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$3

with

$\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$4

for selected neighbors. The fast nearest-Gaussian search RT-GPS bounds error by discarding sufficiently distant Gaussians and uses confidence radii derived from covariance eigenvalues,

$\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$5

The result is a hybrid in which Gaussians supply editability, locality, and physical interaction, while the hash grid and NeRF MLP supply continuous appearance and density prediction (Zieliński et al., 4 Aug 2025).

This design changes the status of Gaussians. In HyRF and compact 3D Gaussian variants, Gaussians remain the render-time primitives. In GENIE, Gaussians are conditioning structure rather than the final composited appearance carrier. That distinction explains both its editing behavior and its performance profile: the paper reports “real-time, locality-aware editing,” but its reported rendering speed is $\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$6 FPS for $\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$7k Gaussians at $\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$8 and $\mathbf{C}=\sum_{i\in I}\mathbf{c}_i\alpha_i\prod_{j=1}^{i-1}(1-\alpha_j).$9 FPS for $\mathbf{x}^a\in\mathbb{R}^3$0M Gaussians at $\mathbf{x}^a\in\mathbb{R}^3$1, so “real-time” refers to immediate representational response without retraining rather than game-rate NeRF rendering (Zieliński et al., 4 Aug 2025).

5. Auxiliary hybrids and adjacent design lines

HAC is the clearest example of a Gaussian–hash-grid construction that is not a persistent hybrid radiance representation. It defines a hash grid

$\mathbf{x}^a\in\mathbb{R}^3$2

queries it at anchor locations,

$\mathbf{x}^a\in\mathbb{R}^3$3

and uses the resulting context feature as a conditional prior for entropy coding:

$\mathbf{x}^a\in\mathbb{R}^3$4

The method predicts adaptive quantization refinements and conditional Gaussian likelihood parameters from $\mathbf{x}^a\in\mathbb{R}^3$5, compresses the binary hash grid itself, and reports over $\mathbf{x}^a\in\mathbb{R}^3$6 size reduction relative to vanilla 3DGS and over $\mathbf{x}^a\in\mathbb{R}^3$7 relative to Scaffold-GS. Crucially, the paper states that no additional operations are required during rendering because the hash grid is removed once $\mathbf{x}^a\in\mathbb{R}^3$8 is fully decoded (Chen et al., 2024).

A broader adjacent literature supplies transferable mechanisms without being Gaussian-primitive hybrids in the strict sense. Zip-NeRF combines Gaussian ray-footprint reasoning with Instant-NGP-style hash-grid encoding; here “Gaussian” denotes finite conical-frustum sample footprints rather than explicit scene primitives, and the key per-level anti-aliased feature is

$\mathbf{x}^a\in\mathbb{R}^3$9

with $\mathcal{A}=\{\mathbf{f}^a\in\mathbb{R}^{D^a},\mathbf{l}\in\mathbb{R}^6,\mathbf{o}\in\mathbb{R}^{3K}\},$0 determined by the footprint size relative to grid cell size (Barron et al., 2023). BiRF demonstrates that a hybrid 2D–3D multi-resolution hash grid can be binarized to $\mathcal{A}=\{\mathbf{f}^a\in\mathbb{R}^{D^a},\mathbf{l}\in\mathbb{R}^6,\mathbf{o}\in\mathbb{R}^{3K}\},$1 while remaining competitive at sub-megabyte storage, but it contains no Gaussian primitives (Shin et al., 2023). GP-NeRF and Hyb-NeRF show two further decompositions: 3D hash grids complemented by high-resolution dense planes for large unbounded scenes, and learnable positional features at coarse levels with hash grids at fine levels (Zhang et al., 2023, Wang et al., 2023). Mip-Grid adds explicit scale-aware filtered grids, arguing that multi-resolution encodings alone do not solve anti-aliasing because the representation must condition on footprint scale, not only on position (Nam et al., 2024).

Structured-explicit methods provide another set of design lessons. GaussianCube turns an unstructured Gaussian set into a fixed $\mathcal{A}=\{\mathbf{f}^a\in\mathbb{R}^{D^a},\mathbf{l}\in\mathbb{R}^6,\mathbf{o}\in\mathbb{R}^{3K}\},$2 tensor by fitting a fixed number of Gaussians and rearranging them into a voxel grid through optimal transport, showing that structured storage of explicit Gaussian parameters can greatly improve downstream learnability even when it does not change reconstruction quality (Zhang et al., 2024). INPC separates an octree-based implicit geometric probability field from a multi-resolution hash-grid appearance field, then extracts an explicit point cloud for differentiable rasterization, demonstrating that geometry and appearance can live in different structures even when the explicit primitive is not a Gaussian (Hahlbohm et al., 2024). GES is not a hash-grid method in its main form, but its Compact-GES variant replaces SH coefficients with hash-grid color lookup, and the overall coarse/fine split—opaque surfels for visibility, Gaussian residuals for detail—is directly relevant to Gaussian–hash-grid hybrids that want coarse explicit geometry plus a learned appearance field (Ye et al., 24 Apr 2025).

6. Empirical behavior, limitations, and design lessons

The empirical record across these papers supports a consistent conclusion: hybridization is most effective when it assigns different representational burdens to explicit primitives and the hash grid rather than attempting a uniform neural replacement of all Gaussian attributes. HyRF reports 8 explicit parameters per Gaussian instead of the 59 parameters of standard 3DGS, plus shared neural-field parameters, and attributes its gains to both smaller per-Gaussian storage and fewer total Gaussians (Wang et al., 21 Sep 2025). Compact 3D Gaussian Representation reports average Mip-NeRF 360 storage dropping from 746 MB in 3DGS to 48.8 MB in the compact model and to 29.1 MB with post-processing, with the color hash remaining the largest component of the compact representation (Lee et al., 2023). Hybrid Latents reduces Mip-NeRF 360 primitive counts from 2.7M in 3DGS, 2.0M in 2DGS, and 1.0M in NeST to 0.7M in its Gaussian variant and 0.2M in its beta-kernel variant, while the sparse beta variant trades some reconstruction quality for much higher efficiency (Kelkar et al., 16 Apr 2026).

The same ablations also mark the limits of the approach. HyRF’s “w/o Explicit” and “w/o Neural” results show that a pure neural-grid prediction of Gaussian attributes loses high-frequency detail, while an isotropic diffuse explicit model loses far more fidelity; the two parts are complementary rather than interchangeable (Wang et al., 21 Sep 2025). Compact 3D Gaussian Representation shows that I-NGP works for color but not for opacity, scale, or rotation, so a hybrid factorization has to be selective (Lee et al., 2023). Hybrid Latents shows that bounded beta kernels and BCE-based sparsification can cut the active primitive set dramatically, but lower reconstruction quality accompanies the most aggressive sparsity regime (Kelkar et al., 16 Apr 2026). GENIE shows that locality-aware Gaussian conditioning yields strong editing behavior, but the rendering loop remains expensive (Zieliński et al., 4 Aug 2025). HAC shows that strong mutual information between unstructured anchors and a structured hash grid can be exploited for compression, but also that such mutual information does not by itself imply a useful persistent render-time hybrid (Chen et al., 2024).

Three design lessons recur across the corpus. First, a successful Gaussian–hash-grid hybrid usually preserves an explicit residual pathway on the primitives; fully neuralized Gaussian attributes are repeatedly described as too low-pass or too unstable for high-frequency structure (Wang et al., 21 Sep 2025, Kelkar et al., 16 Apr 2026). Second, appearance is the most reliable target for hash-grid sharing: view-dependent color replacement works well, whereas geometry, opacity, and anisotropy often remain better behaved when at least partly explicit (Lee et al., 2023). Third, the word “hybrid” should be interpreted literally: the most effective systems use the hash grid to do something the Gaussian substrate does poorly—shared radiance, residual texture, conditional priors, or editable latent support—rather than duplicating the explicit representation.