K-Planes: Factorized Radiance Field Model

• K-Planes is a radiance field model that factorizes a d-dimensional scene into exactly (d choose 2) learnable 2D feature planes, handling both static and dynamic cases.
• It employs bilinear interpolation and multiplicative (Hadamard) fusion to combine features, with both an explicit linear decoder and a hybrid MLP option for decoding.
• The model delivers significant memory compression and interpretability while enabling dimension-specific priors like temporal smoothness and multi-resolution spatial regularization.

Searching arXiv for the K-Planes radiance-field literature and close extensions. K-Planes is a white-box model for radiance fields in arbitrary dimensions that represents a $d$-dimensional scene using $\binom d2$ learnable feature planes, one for each unordered pair of dimensions. It was introduced as a planar factorization that supports a seamless transition from static scenes ($d=3$) to dynamic scenes ($d=4$), while making dimension-specific priors such as temporal smoothness and multi-resolution spatial structure straightforward to impose (Fridovich-Keil et al., 2023). In the original formulation, features sampled from the planes are fused by element-wise multiplication and decoded either by a linear feature decoder with a learned color basis or by a hybrid MLP decoder; the method is reported to achieve competitive and often state-of-the-art reconstruction fidelity with low memory usage, including approximately $1000\times$ compression over a full $4$D grid (Fridovich-Keil et al., 2023).

1. Definition and representational form

K-Planes represents a $d$-dimensional scene with exactly $k=\binom d2$ feature planes, each corresponding to one unordered pair of coordinates. The canonical examples given are tri-planes for $d=3$, namely $(xy,xz,yz)$, hex-planes for $\binom d2$0, namely $\binom d2$1, and deca-planes for $\binom d2$2 (Fridovich-Keil et al., 2023). Each plane $\binom d2$3 is a learnable $\binom d2$4D grid

$\binom d2$5

where $\binom d2$6 is the plane resolution and $\binom d2$7 is the per-plane feature dimension (Fridovich-Keil et al., 2023).

A query point $\binom d2$8 is projected to each plane, bilinearly interpolated, and then fused across all planes. Using the notation of the original formulation,

$\binom d2$9

with $d=3$0 denoting bilinear interpolation, and the fused feature is

$d=3$1

The fusion is explicitly Hadamard, rather than additive (Fridovich-Keil et al., 2023).

The representation is also described as naturally extensible across dimensions. This suggests that the central abstraction is not tied to conventional $d=3$2D radiance fields alone, but to axis-paired factorization of higher-dimensional scene variables. A plausible implication is that K-Planes is best understood as a coordinate-factorized feature field whose inductive bias is imposed directly through the geometry of the planes rather than through a deep implicit decoder.

2. Querying, decoding, and volumetric rendering

The original model supports two decoder families. In the fully explicit linear decoder, a learned color basis is produced by a small MLP $d=3$3 conditioned on view direction $d=3$4, yielding basis vectors $d=3$5, while density uses a fixed learned basis $d=3$6 (Fridovich-Keil et al., 2023). The decoded density and color are

$d=3$7

The density is mapped through $d=3$8 with clamped gradient to enforce nonnegativity, and color through $d=3$9 to enforce $d=4$0 (Fridovich-Keil et al., 2023).

The hybrid decoder instead uses

$d=4$1

where $d=4$2 and $d=4$3 (Fridovich-Keil et al., 2023).

Rendering follows standard volumetric rendering:

$d=4$4

Training uses a photometric square loss

$d=4$5

with optional regularization terms for spatial total variation, temporal smoothness, and static-dynamic separation (Fridovich-Keil et al., 2023).

The explicit decoder is described as white-box, fast, and small, whereas the hybrid decoder uses a small black-box MLP and is harder to interpret (Fridovich-Keil et al., 2023). Within the paper’s terminology, the interpretability claim is tied to the fact that the model is entirely linear after selecting $d=4$6.

3. Dimension-specific priors and static-dynamic decomposition

A central feature of K-Planes is that the planar factorization makes it easy to impose priors on specific subsets of planes. Multi-resolution spatial structure is implemented by replicating planes at several spatial scales and concatenating their features; lower-resolution planes capture coarse structure, while higher-resolution planes add detail (Fridovich-Keil et al., 2023).

For space-only planes $d=4$7, the paper introduces a spatial total-variation penalty

$d=4$8

For space-time planes $d=4$9, temporal smoothness is imposed through a second-order penalty along the time dimension,

$1000\times$0

To encourage sparse transients, the space-time planes are initialized to the all-ones matrix and regularized with

$1000\times$1

At inference, setting the three time-planes to $1000\times$2 yields a purely static rendering, with the residual interpreted as the dynamic component (Fridovich-Keil et al., 2023).

This decomposition is one of the main conceptual differences between K-Planes and radiance-field parameterizations that treat all dimensions uniformly inside a single black-box network. The paper explicitly presents this as a natural decomposition of static and dynamic components of a scene (Fridovich-Keil et al., 2023).

4. Efficiency, compression, and reported empirical behavior

The original K-Planes paper reports that a $1000\times$3D dense grid at $1000\times$4 frames would be larger than $1000\times$5 GB, whereas K-Planes uses six $1000\times$6D grids, plus multiscale copies, totaling at most $1000\times$7 MB, which is summarized as approximately $1000\times$8 compression (Fridovich-Keil et al., 2023). The implementation is described as pure PyTorch, without custom CUDA, with typical optimization settings using Adam and a learning-rate decay from $1000\times$9 to $4$0 over roughly $4$1k to $4$2k steps (Fridovich-Keil et al., 2023).

Reported single-GPU training times on an A30 span several scene classes:

Scene class	Training time	Parameters
Static synthetic NeRF scenes	$4$3 min	$4$4 M
LLFF real forward scenes	$4$5 min	$4$6 M
D-NeRF synthetic dynamic	$4$7 min	$4$8 M
DyNeRF real dynamic	$4$9–$d$0 hrs	$d$1–$d$2 M
Phototourism	$d$3 min	$d$4 M

The same work reports mean reconstruction quality on standard benchmarks as $d$5 dB / $d$6 for static synthetic NeRF scenes, $d$7 dB / $d$8 for LLFF, $d$9 dB / $k=\binom d2$0 for D-NeRF, $k=\binom d2$1 dB / $k=\binom d2$2 for DyNeRF, and $k=\binom d2$3 dB / $k=\binom d2$4 MS-SSIM for Phototourism (Fridovich-Keil et al., 2023). The abstract characterizes this performance as competitive and often state-of-the-art, with low memory usage (Fridovich-Keil et al., 2023).

The ablation results given in the paper are tightly tied to specific modeling choices. Removing Hadamard fusion produces a PSNR drop of approximately $k=\binom d2$5 dB in the explicit model, turning off time smoothness yields a drop of approximately $k=\binom d2$6 dB on dynamic sequences, and varying the feature length $k=\binom d2$7 from $k=\binom d2$8 to $k=\binom d2$9 trades PSNR from $d=3$0 to $d=3$1 on Lego (Fridovich-Keil et al., 2023). These observations are used to justify the explicit model’s reliance on multiplicative plane fusion.

5. Extensions: TK-Planes and UAV-based dynamic scenes

An explicit extension, "TK-Planes: Tiered K-Planes with High Dimensional Feature Vectors for Dynamic UAV-based Scenes" (Maxey et al., 2024), adapts the tri-plane form to dynamic UAV-based perception by introducing tiered feature vectors, an image decoder, and an explicit separation between static and dynamic feature planes. In the description provided, standard K-Planes in the $d=3$2D case uses three orthogonal $d=3$3D feature grids

$d=3$4

associated with $d=3$5, $d=3$6, and $d=3$7, sampled as

$d=3$8

and combined by

$d=3$9

A small MLP predicts density and color, and rendering follows the standard NeRF quadrature (Maxey et al., 2024).

TK-Planes generalizes this structure by introducing $(xy,xz,yz)$0 tiers and, at each tier, nine $(xy,xz,yz)$1D planes of $(xy,xz,yz)$2-dimensional features: three static spatial planes, three dynamic spatial planes, and three dynamic spatio-temporal planes (Maxey et al., 2024). The $(xy,xz,yz)$3-th tier static feature is

$(xy,xz,yz)$4

while the dynamic feature multiplies both dynamic spatial and dynamic spatio-temporal factors:

$(xy,xz,yz)$5

The full feature is then

$(xy,xz,yz)$6

Per-tier rendered feature maps are decoded by a U-Net-style convolutional decoder, with losses comprising reconstruction, a cosine-based static/dynamic separation term, and an optional feature smoothness regularizer (Maxey et al., 2024).

The abstract of TK-Planes states that the method is designed to bridge the domain gap between synthetic and real-world data for UAV-based perception in dynamic scenes with small moving objects or human actions, and that it observes considerable improvement in accuracy over state-of-the-art neural rendering methods on Okutama Action and UG2 (Maxey et al., 2024). Because the provided details do not include numerical metrics, more specific comparative claims would require consulting the paper directly.

6. Scope, terminology, and distinctions from other uses of “k-plane”

In the radiance-field literature, "K-Planes" refers to the specific factorized representation introduced for explicit radiance fields in space, time, and appearance (Fridovich-Keil et al., 2023). This usage is distinct from several established meanings of "k-plane" in mathematics and theoretical computer science.

In differential geometry and convex analysis, a real or complex $(xy,xz,yz)$7-plane is a $(xy,xz,yz)$8-dimensional affine subspace, and such terminology appears in results on convex hypersurfaces and the Levi form (Nikolov, 2012). In integral geometry, $(xy,xz,yz)$9-planes are the integration domains of classical and mixed Radon transforms on affine Grassmannians (Rubin et al., 2019, Salman, 2018). In geometric combinatorics and finite geometry, $\binom d2$00-planes denote affine or quadratic subspaces over finite fields, with associated incidence and orthogonality graphs (Phuong et al., 2015, Yoo, 2020). In graph drawing and topological graph theory, a $\binom d2$01-plane graph is a graph drawing in which every edge is crossed at most $\binom d2$02 times (Hummel et al., 2019, Bekos et al., 2018).

This terminological overlap can produce a superficial ambiguity. However, the capitalized form "K-Planes" in contemporary computer vision refers specifically to the radiance-field model of Fridovich-Keil and colleagues (Fridovich-Keil et al., 2023). A plausible implication is that the capitalization is not merely stylistic but functions as a disambiguating proper name for the representation.

7. Significance and limitations

The contribution of K-Planes is framed around three linked properties: explicitness, factorized dimensionality, and decoder simplicity. The paper’s summary lists as main contributions a white-box factorization of any $\binom d2$03-dimensional radiance field into exactly $\binom d2$04 planes, a natural extension from static to dynamic and higher-dimensional volumes, straightforward dimension-specific priors, a linear decoder with a learned color basis, and extreme compression together with fast single-GPU training in PyTorch (Fridovich-Keil et al., 2023).

The stated strengths are interpretability, flexibility across static, dynamic, and variable-appearance scenes, efficiency from a small memory footprint and lack of custom kernels, and competitive or better PSNR/SSIM relative to recent MLP-based or hybrid methods (Fridovich-Keil et al., 2023). The stated limitations are equally specific: the method still relies on a small MLP for the learned color basis; pure linear decoding with spherical harmonics was reported as less expressive; multiplicative fusion can suffer from underflow if features become too small; and high-frequency effects beyond the learned basis capacity may require more basis dimensions or a larger MLP (Fridovich-Keil et al., 2023).

The future directions named in the paper include learned plane rotations or non-axis-aligned planes, extensions to $\binom d2$05D and higher-dimensional appearance or transient latent spaces, integration with learned scene priors for few-shot or single-frame dynamic reconstruction, and more elaborate fusion functions such as gated products or attention (Fridovich-Keil et al., 2023). These proposals indicate that K-Planes is both a concrete radiance-field architecture and a broader design pattern: representing a high-dimensional scene by low-dimensional coordinate-pair factors whose interaction is made explicit.