Voxelized Radiance Field (VoxelRF)

• Voxelized Radiance Field (VoxelRF) is a radiance field method that integrates explicit voxel structures into scene representation, enabling faster rendering and efficient optimization.
• It combines diverse forms such as dense grids, sparse voxel octrees, and hybrid voxel conditioning to enhance geometry priors and guide sampling in volume rendering.
• VoxelRF methods balance trade-offs between training speed, memory constraints, and fine detail accuracy, with applications spanning object reconstruction and wireless modeling.

Searching arXiv for the cited VoxelRF-related papers to ground the article in current arXiv records. Voxelized radiance field (VoxelRF) denotes a family of radiance-field methods that introduce explicit voxelized spatial structure into scene representation, rendering, or optimization. Across the cited literature, this structure may be the primary scene parameterization—dense grids, sparse voxel grids, sparse voxel octrees, spherical voxels, or factorized voxel tensors—or it may function as a geometric prior, a conditioning scaffold, a sampling guide, or a regularization domain while rendering remains NeRF-like and continuous (Sun et al., 2022, Li et al., 2022, Aroudj et al., 2022, Yu et al., 2022). The unifying idea is the relocation of a substantial part of scene organization from a large coordinate MLP into a spatially indexed 3D lattice or hierarchy, typically queried by interpolation and coupled to differentiable volume accumulation.

1. Conceptual scope and taxonomy

The literature does not treat VoxelRF as a single architecture. One branch is explicitly voxel-native. DVGOv2 uses the “simplest dense grid representation”; ReVoRF inherits a DVGO-style voxel grid storing density and RGB-related features; OmniVoxel replaces a fully implicit omnidirectional NeRF with spherical feature voxels; DiffRF directly generates dense explicit voxel radiance fields; RGB-D Plenoxel mapping uses an explicit sparse voxel grid with density and spherical harmonics; ERF uses a sparse voxel octree with an opacity field and a surface light field; wireless VoxelRF and Voxel-CKM both move wireless radiance-style quantities into voxel grids (Sun et al., 2022, Xu et al., 2024, Li et al., 2022, Müller et al., 2022, Teigen et al., 2023, Aroudj et al., 2022, Zeng et al., 14 Jul 2025, Li et al., 2 Jun 2026).

A second branch is hybrid. PVSeRF conditions a continuous radiance field on a predicted $32^3$ volumetric feature grid and a point-cloud branch; VoxNeRF uses a Sparse Voxel Octree geometry prior to guide sampling while the learned radiance remains hash-grid-plus-MLP; ShaRF uses a $128^3$ occupancy scaffold to condition a NeRF-like radiance MLP; V4D stores latent voxel features and decodes them with small MLPs; CVT-xRF keeps the scene implicit but uses voxels as a training structure for ray grouping, local neighborhood formation, and contrastive regularization (Yu et al., 2022, Wang et al., 2023, Rematas et al., 2021, Gan et al., 2022, Zhong et al., 2024).

Class	Voxel role	Representative papers
Explicit voxel RF	Scene parameters stored directly in voxelized form	DVGOv2, ReVoRF, OmniVoxel, DiffRF, ERF, RGB-D Plenoxel, VoxelRF, Voxel-CKM
Hybrid voxel-conditioned RF	Voxels condition, guide, or regularize a continuous field	PVSeRF, VoxNeRF, ShaRF, V4D, CVT-xRF

A recurrent misconception is therefore that “voxelized radiance field” must mean direct storage of final density and color at every voxel. The cited papers show a broader usage: voxels can be the radiance field itself, but they can also be the geometric substrate through which a radiance field is conditioned or trained (Yu et al., 2022, Wang et al., 2023, Zhong et al., 2024).

2. Representational forms

The most direct formulations store density and appearance-related quantities in regular grids. ReVoRF describes voxelized feature representations that “store the RGB features $f_c$ and density $\sigma$ in a voxel grids,” and wireless VoxelRF uses a density voxel grid together with a feature voxel grid queried by trilinear interpolation (Xu et al., 2024, Zeng et al., 14 Jul 2025). RGB-D Plenoxel mapping is more explicit still: each sparse-grid vertex stores 28 scalar parameters, comprising one density value and 27 spherical-harmonics color parameters (Teigen et al., 2023). DiffRF operates on a dense $32^3$ voxel tensor representing a pre-activated field of density and color channels, which makes the radiance field convolution-friendly and directly compatible with volumetric denoising diffusion (Müller et al., 2022).

Other methods alter the voxel parameterization to match imaging geometry or scene scale. OmniVoxel adopts spherical voxelization along $(r,\theta,\phi)$ rather than Cartesian $(x,y,z)$ and combines it with a TensoRF-style vector-matrix decomposition of spherical feature tensors (Li et al., 2022). ERF replaces dense grids with a sparse voxel octree in which every node, including inner nodes, stores opacity and spherical-harmonics surface-radiance parameters, yielding a multi-resolution explicit field queried jointly in space and LoD (Aroudj et al., 2022). Voxel-CKM uses a plane-line vector-matrix decomposition for RF voxel fields and gives storage scaling of $O\big((R_1+R_2)N^2+R_2D\big)$ rather than dense-grid $O(N^3D)$ (Li et al., 2 Jun 2026).

Hybrid methods use voxels as structured priors rather than final outputs. PVSeRF reconstructs a volumetric feature grid $$\mathbf F_V\in\mathbb R^{32\times 32\times 32\times C}$$ from a single image and samples voxel-aligned features through multi-scale trilinear interpolation, while a separate 1024-point cloud provides complementary surface-aligned detail (Yu et al., 2022). VoxNeRF obtains geometry with MonoSDF, converts it into a Sparse Voxel Octree, and uses the SVO only for ray-surface intersections and sampling; the actual learned radiance remains a hash-grid representation (Wang et al., 2023). ShaRF similarly confines voxels to occupancy, using a generated $128^3$0 scaffold as an explicit geometry prior for a continuous radiance network (Rematas et al., 2021).

This diversity suggests that VoxelRF is better understood as a representational axis than as a single model family: the decisive question is not whether a method contains voxels, but whether voxelization materially organizes radiance, geometry, or optimization.

3. Rendering and optimization principles

Despite representational variation, the dominant rendering pattern is differentiable volume accumulation. A representative form is

$128^3$1

with transmittance determined by the accumulated density of earlier samples. This form appears, with small paper-specific notational differences, in OmniVoxel, V4D, RGB-D Plenoxel mapping, wireless VoxelRF, and related methods (Li et al., 2022, Gan et al., 2022, Teigen et al., 2023, Zeng et al., 14 Jul 2025). ReVoRF writes the continuous rendering equation in standard NeRF integral form and then builds its sparse-view training scheme on top of that voxel backbone (Xu et al., 2024).

The optimization landscape differs sharply from MLP-centric NeRF. DVGOv2 emphasizes direct optimization of dense voxel grids, adds fused CUDA kernels, early ray termination, and an $128^3$2 reformulation of the mip-NeRF 360 distortion loss, and reports a 2–3x speedup over DVGO (Sun et al., 2022). ERF goes further toward explicitness: it reconstructs a sparse voxel octree “from scratch” by stochastic gradient descent with an inverse differentiable renderer, and it uses opacity compositing rather than standard density-based Beer–Lambert transmittance (Aroudj et al., 2022).

Several papers make voxel-specific regularization central rather than auxiliary. ReVoRF introduces bilateral geometric consistency, reliability-guided learning, and reliability-aware voxel smoothing to use unreliable pseudo-views instead of discarding them (Xu et al., 2024). CVT-xRF uses voxel partitions to ensure sampled rays intersect a common local region, then imposes in-voxel consistency through a Transformer and a contrastive loss (Zhong et al., 2024). RGB-D Plenoxel mapping derives analytical gradients of rendered RGB and rendered expected depth with respect to density, color, samples, rays, and pose, enabling direct mapping and tracking without a neural network (Teigen et al., 2023).

A further distinction concerns whether rendering is purely optical. Voxel-CKM and wireless VoxelRF reinterpret volume accumulation for radio propagation: the voxel field stores RF-specific quantities, and rendering aggregates direction-conditioned or complex RF contributions along rays rather than optical RGB radiance (Li et al., 2 Jun 2026, Zeng et al., 14 Jul 2025). This extends the VoxelRF concept from image-based view synthesis to spatial wireless field prediction.

4. Problem settings and application domains

Voxelized radiance fields appear in markedly different regimes. In single-image object-centric reconstruction, PVSeRF addresses feature ambiguity in pixel-aligned radiance fields by adding voxel-aligned and surface-aligned geometric conditioning, and ShaRF uses a voxel occupancy scaffold to disentangle shape and appearance under single-view inversion (Yu et al., 2022, Rematas et al., 2021). In sparse-view scene reconstruction, ReVoRF specializes a voxel backbone for 3–4 posed inputs, and CVT-xRF introduces voxel-aware 3D consistency learning for sparse-input NeRF training (Xu et al., 2024, Zhong et al., 2024).

Indoor sparse-view synthesis motivates another line. VoxNeRF introduces a geometry prior in the form of an SVO derived from MonoSDF and uses voxel-guided sampling to concentrate computation near probable surfaces in ScanNet-like settings (Wang et al., 2023). Omnidirectional capture motivates OmniVoxel, whose spherical voxelization is designed for equirectangular panoramas and unbounded 360° scenes (Li et al., 2022). Dynamic scenes motivate V4D, which stores two $128^3$3 voxel volumes and conditions tiny MLP decoders on time for 4D novel-view synthesis (Gan et al., 2022).

Generative modeling forms a distinct strand. DiffRF operates directly in voxel-radiance-field space with a 3D U-Net denoiser and supplements voxel-space diffusion with a rendering-guided objective so that the learned prior prefers good rendered images rather than imitating fitting artifacts (Müller et al., 2022). Mapping and tracking form another strand: RGB-D Plenoxel mapping uses an explicit radiance-field map as a dense SLAM-like backend, arguing that depth is essential when geometry, rather than only view synthesis, is the goal (Teigen et al., 2023).

Wireless modeling broadens the term beyond optical radiance. Wireless VoxelRF uses voxelized density and feature grids plus two shallow MLPs to synthesize spatial spectra for a moving transmitter and fixed receiver, while Voxel-CKM formulates channel knowledge map construction as an RF analogue of an explicit factorized voxel radiance field (Zeng et al., 14 Jul 2025, Li et al., 2 Jun 2026). This suggests that “radiance field” in the VoxelRF literature has already expanded into a more general volumetric signal-field paradigm.

5. Empirical characteristics

A primary empirical theme is acceleration. OmniVoxel reports reducing omnidirectional scene training from roughly $128^3$4 hours to $128^3$5 minutes per scene (Li et al., 2022). VoxNeRF reports 5 minutes training and 0.5 s rendering time, compared with 15 minutes and 0.66 s for Instant-NGP in its reported setup (Wang et al., 2023). ReVoRF reports 3 FPS rendering, 7 minutes to train a $128^3$6 scene, and 11 minutes on LLFF in the 3-view setting (Xu et al., 2024). Wireless VoxelRF reports 20 minutes training and 0.042 s inference, versus 15h30m and 0.396 s for NeRF$128^3$7 on its RFID benchmark (Zeng et al., 14 Jul 2025). DiffRF shows the opposite split: the explicit voxel field renders at over 380 FPS after synthesis, but 1000-step DDPM sampling still costs 48.6 seconds per sample (Müller et al., 2022).

Speed, however, is not the only pattern. Voxelization is frequently tied to better geometry or stronger data efficiency. PVSeRF improves on pixelNeRF in category-agnostic ShapeNet metrics by combining pixel-, voxel-, and surface-aligned features, with mean PSNR 27.48 versus 26.80 and LPIPS 0.096 versus 0.108 (Yu et al., 2022). ReVoRF’s ablation shows a progression from 17.19 PSNR for the DVGO baseline to 20.72 after adding reliability-aware components, with the largest late gain attributed to reliability-aware voxel smoothing (Xu et al., 2024). CVT-xRF raises NeRF on DTU 3-view from 6.68 to 14.13 PSNR and improves BARF and SPARF as well, supporting the claim that voxel-local regularization mitigates sparse-view ambiguity (Zhong et al., 2024).

The main counterweight is memory and resolution scaling. Dense voxel grids remain cubic in resolution unless compressed or factorized. DVGOv2 explicitly notes that dense grids limit achievable resolution on difficult unbounded scenes (Sun et al., 2022). V4D reports 377 MB versus 13 MB for D-NeRF, illustrating the classic VoxelRF tradeoff of capacity and speed against storage cost (Gan et al., 2022). ReVoRF explicitly notes a tendency toward smoothed results and loss of fine details under dense-grid constraints (Xu et al., 2024). Factorized methods such as OmniVoxel and Voxel-CKM are responses to this structural issue rather than exceptions to it (Li et al., 2022, Li et al., 2 Jun 2026).

A second misconception concerns geometry. RGB-D Plenoxel mapping shows that RGB-only optimization can achieve higher PSNR yet much worse geometry: on a reported Replica example, RGB-only gives PSNR 30.612 with average L1 depth error about 0.6971 m/pixel, whereas RGB-D training gives PSNR 28.570 with depth error about 0.0090 m/pixel (Teigen et al., 2023). The result is a direct warning against equating visually plausible renderings with accurate 3D structure.

6. Limitations, boundary cases, and bibliographic cautions

The principal limitation is that voxelization does not remove the need to choose what voxels mean. In explicit methods, voxels may store density, opacity, RGB features, spherical-harmonics coefficients, latent features, or decomposed plane-line factors; in hybrid methods, they may store only occupancy or only a geometry prior (Aroudj et al., 2022, Teigen et al., 2023, Yu et al., 2022, Rematas et al., 2021). Consequently, papers that share the VoxelRF label can differ substantially in rendering semantics, learnable parameters, and failure modes.

Dense or semi-dense grids also inherit bounded-scene and memory-resolution assumptions. ReVoRF remains constrained by dense-grid smoothness and sparse pseudo-view uncertainty; VoxNeRF depends on the quality of the geometry prior produced by MonoSDF or SLAM; ShaRF and PVSeRF are object-centric and single-image conditioned; V4D gains speed over large MLPs but still incurs high memory; DiffRF is restricted by its dense $128^3$8 field and view-independent final representation (Xu et al., 2024, Wang et al., 2023, Rematas et al., 2021, Yu et al., 2022, Gan et al., 2022, Müller et al., 2022).

The literature also exposes a conceptual boundary case: not every voxel-aware radiance-field method is voxel-native. CVT-xRF is best read as a voxel-aware regularization framework rather than a radiance field stored in voxels, and VoxNeRF is better described as voxel-guided than as a pure voxel RF (Zhong et al., 2024, Wang et al., 2023). Any taxonomy of VoxelRF therefore needs to distinguish representation-level voxelization from training-structure voxelization.

A bibliographic caution concerns “PERF: Performant, Explicit Radiance Fields” (Rasmuson et al., 2021). The supplied content for that entry was reported as an IEEEtran template with no technical material about PERF, radiance fields, voxels, or experiments. No technical characterization of PERF can therefore be supported from that source alone (Rasmuson et al., 2021).

Taken together, these works position VoxelRF not as a single method but as an explicit-structure program in radiance-field research: replace part of the scene-function burden carried by large coordinate networks with spatially indexed voxelized organization, then exploit that organization for faster optimization, stronger geometry priors, structured regularization, or domain-specific rendering. The resulting family spans dense-grid optimization, sparse octrees, spherical and factorized tensors, hybrid voxel-conditioned NeRFs, diffusion over voxel fields, RGB-D tracking backends, and RF channel-field models (Sun et al., 2022, Aroudj et al., 2022, Müller et al., 2022, Zeng et al., 14 Jul 2025).