---
title: 'K-Planes: Factorized Radiance Field Model'
url: https://www.emergentmind.com/topics/k-planes
type: topic
---

# K-Planes: Factorized Radiance Field Model

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 [2301.10241]. 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 [2301.10241].

## 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 $d=4$, namely $(xy,xz,yz,xt,yt,zt)$, and deca-planes for $d=5$ [2301.10241]. Each plane $c\in C$ is a learnable $2$D grid
$$
P_{(c)} \in \mathbb{R}^{N\times N\times M},
$$
where $N$ is the plane resolution and $M$ is the per-plane feature dimension [2301.10241].

A query point $q\in\mathbb{R}^d$ is projected to each plane, bilinearly interpolated, and then fused across all planes. Using the notation of the original formulation,
$$
f_{(c)}(q)=\psi\bigl(P_{(c)},\pi_{(c)}(q)\bigr),
$$
with $\psi$ denoting bilinear interpolation, and the fused feature is
$$
f(q)=\prod_{c\in C} f_{(c)}(q)\in\mathbb{R}^M.
$$
The fusion is explicitly Hadamard, rather than additive [2301.10241].

The representation is also described as naturally extensible across dimensions. This suggests that the central abstraction is not tied to conventional $3$D 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 $b_{\mathrm{color}}(\cdot)$ conditioned on view direction $d$, yielding basis vectors $b_R(d),b_G(d),b_B(d)\in\mathbb{R}^M$, while density uses a fixed learned basis $b_\sigma\in\mathbb{R}^M$ [2301.10241]. The decoded density and color are
$$
\sigma(q)=f\cdot b_\sigma,\qquad
c(q,d)=\bigl[f\cdot b_R(d),\,f\cdot b_G(d),\,f\cdot b_B(d)\bigr].
$$
The density is mapped through $\exp(\sigma)$ with clamped gradient to enforce nonnegativity, and color through $\mathrm{sigmoid}(c)$ to enforce $[0,1]^3$ [2301.10241].

The hybrid decoder instead uses
$$
\sigma(q),\hat f(q)=g_\sigma\bigl(f(q)\bigr),\quad
c(q,d)=g_{RGB}\bigl(\hat f(q),\gamma(d)\bigr),
$$
where $g_\sigma:\mathbb{R}^M\to(\mathbb{R},\mathbb{R}^{\hat M})$ and $g_{RGB}:\mathbb{R}^{\hat M}\times\mathbb{R}^{D_d}\to\mathbb{R}^3$ [2301.10241].

Rendering follows standard volumetric rendering:
$$
\hat C(r)=\sum_i T_i\bigl(1-e^{-\sigma_i\Delta_i}\bigr)c_i,\quad
T_i=\exp\!\Bigl(-\sum_{j<i}\sigma_j\Delta_j\Bigr).
$$
Training uses a photometric square loss
$$
\mathcal L_{photo}=\sum_{rays}\Big\|\hat C(r)-C_{gt}(r)\Big\|_2^2,
$$
with optional regularization terms for spatial total variation, temporal smoothness, and static-dynamic separation [2301.10241].

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 [2301.10241]. Within the paper’s terminology, the interpretability claim is tied to the fact that the model is entirely linear after selecting $f(q)$.

## 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 [2301.10241].

For space-only planes $c\in\{xy,xz,yz\}$, the paper introduces a spatial total-variation penalty
$$
\mathcal L_{TV}
= \frac{1}{|C|\,N^2}\sum_{c,i,j}
\bigl\|P_c^{i,j}-P_c^{i-1,j}\bigr\|_2^2
+\bigl\|P_c^{i,j}-P_c^{i,j-1}\bigr\|_2^2.
$$
For space-time planes $c\in\{xt,yt,zt\}$, temporal smoothness is imposed through a second-order penalty along the time dimension,
$$
\mathcal L_{time}
= \frac{1}{|C|\,N^2}\sum_{c,i,t}
\bigl\|P_c^{i,t-1}-2P_c^{i,t}+P_c^{i,t+1}\bigr\|_2^2.
$$
To encourage sparse transients, the space-time planes are initialized to the all-ones matrix and regularized with
$$
\mathcal L_{sep}
= \sum_{c\in\{xt,yt,zt\}}\|\mathbf 1-P_c\|_1.
$$
At inference, setting the three time-planes to $1$ yields a purely static rendering, with the residual interpreted as the dynamic component [2301.10241].

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 [2301.10241].

## 4. Efficiency, compression, and reported empirical behavior

The original K-Planes paper reports that a $4$D dense grid at $512^3\times T$ frames would be larger than $300$ GB, whereas K-Planes uses six $2$D grids, plus multiscale copies, totaling at most $200$ MB, which is summarized as approximately $1000\times$ compression [2301.10241]. The implementation is described as pure PyTorch, without custom CUDA, with typical optimization settings using Adam and a learning-rate decay from $5\cdot10^{-3}$ to $5\cdot10^{-4}$ over roughly $40$k to $200$k steps [2301.10241].

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

| Scene class | Training time | Parameters |
|---|---:|---:|
| Static synthetic NeRF scenes | $\sim 38$ min | $32$ M |
| LLFF real forward scenes | $\sim 33$ min | $19$ M |
| D-NeRF synthetic dynamic | $\sim 52$ min | $37$ M |
| DyNeRF real dynamic | $\sim 1.8$–$3.7$ hrs | $27$–$51$ M |
| Phototourism | $\sim 35$ min | $36$ M |

The same work reports mean reconstruction quality on standard benchmarks as $32.2$ dB / $0.960$ for static synthetic NeRF scenes, $26.8$ dB / $0.841$ for LLFF, $31.1$ dB / $0.97$ for D-NeRF, $30.9$ dB / $0.960$ for DyNeRF, and $22.3$ dB / $0.859$ MS-SSIM for Phototourism [2301.10241]. The abstract characterizes this performance as competitive and often state-of-the-art, with low memory usage [2301.10241].

The ablation results given in the paper are tightly tied to specific modeling choices. Removing Hadamard fusion produces a PSNR drop of approximately $6$ dB in the explicit model, turning off time smoothness yields a drop of approximately $1$ dB on dynamic sequences, and varying the feature length $M$ from $2$ to $128$ trades PSNR from $30.6$ to $35.4$ on Lego [2301.10241]. 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" [2405.02762], 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 $3$D case uses three orthogonal $2$D feature grids
$$
F_1:\mathbb{R}^2\to\mathbb{R}^c,\quad
F_2:\mathbb{R}^2\to\mathbb{R}^c,\quad
F_3:\mathbb{R}^2\to\mathbb{R}^c,
$$
associated with $(xy)$, $(yz)$, and $(zx)$, sampled as
$$
f_1=F_1(x,y),\quad f_2=F_2(y,z),\quad f_3=F_3(z,x),
$$
and combined by
$$
f(q)=f_1\odot f_2\odot f_3.
$$
A small MLP predicts density and color, and rendering follows the standard NeRF quadrature [2405.02762].

TK-Planes generalizes this structure by introducing $K$ tiers and, at each tier, nine $2$D planes of $D_k$-dimensional features: three static spatial planes, three dynamic spatial planes, and three dynamic spatio-temporal planes [2405.02762]. The $k$-th tier static feature is
$$
f_s^{(k)}(q)=F^{(k)}_{s,xy}(x,y)\odot F^{(k)}_{s,yz}(y,z)\odot F^{(k)}_{s,zx}(z,x),
$$
while the dynamic feature multiplies both dynamic spatial and dynamic spatio-temporal factors:
$$
f_d^{(k)}(q,\tau)=
\bigl[F^{(k)}_{d,xy}(x,y)\odot F^{(k)}_{d,yz}(y,z)\odot F^{(k)}_{d,zx}(z,x)\bigr]
\odot
\bigl[F^{(k)}_{t,xy}(x,y,\tau)\odot F^{(k)}_{t,yz}(y,z,\tau)\odot F^{(k)}_{t,zx}(z,x,\tau)\bigr].
$$
The full feature is then
$$
f^{(k)}(q,\tau)=f_s^{(k)}(q)\oplus f_d^{(k)}(q,\tau)\in\mathbb{R}^{2D_k}.
$$
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 [2405.02762].

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 [2405.02762]. 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 [2301.10241]. 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 $k$-plane is a $k$-dimensional affine subspace, and such terminology appears in results on convex hypersurfaces and the Levi form [1208.4676]. In integral geometry, $k$-planes are the integration domains of classical and mixed Radon transforms on affine Grassmannians [1901.01150; 1801.05838]. In geometric combinatorics and finite geometry, $k$-planes denote affine or quadratic subspaces over finite fields, with associated incidence and orthogonality graphs [1510.03481; 2004.10742]. In graph drawing and topological graph theory, a $k$-plane graph is a graph drawing in which every edge is crossed at most $k$ times [1908.08905; 1802.10300].

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 [2301.10241]. 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 $d$-dimensional radiance field into exactly $k=\binom d2$ 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 [2301.10241].

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 [2301.10241]. 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 [2301.10241].

The future directions named in the paper include learned plane rotations or non-axis-aligned planes, extensions to $5$D 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 [2301.10241]. 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.

Source: https://www.emergentmind.com/topics/k-planes