Tri-Plane 3D Feature Representation

• Tri-plane volumetric feature representation is a factorized 3D approach that stores features on three orthogonal 2D planes for robust, efficient querying and decoding.
• It enables efficient neural rendering, implicit geometry, and compression by decoupling explicit spatial structure from lightweight implicit decoding.
• Recent variants extend the model with multi-scale, cylindrical, and hybrid coordinate systems to address feature entanglement and improve full-view consistency.

Searching arXiv for recent and foundational papers on tri-plane volumetric feature representation and closely related variants. Tri-plane volumetric feature representation is a factorized 3D parameterization in which volumetric information is stored on three orthogonal 2D feature planes, typically aligned with the $XY$, $YZ$, and $XZ$ axes. A 3D query point is projected onto each plane, plane features are retrieved—most commonly by bilinear interpolation—and the resulting features are aggregated and decoded into density, color, signed distance, occupancy, or other task-specific quantities. In the recent literature, this representation has become a common interface between explicit spatial structure and lightweight implicit decoding in 3D-aware image generation, neural rendering, surface reconstruction, SLAM, segmentation, compression, and robotics, largely because it replaces cubic-memory volumetric grids with a plane factorization that preserves fast continuous querying (Jia et al., 21 Jul 2025, Wang et al., 2023, Cui et al., 2024).

1. Formal definition and feature querying

In its standard Cartesian form, a tri-plane consists of three learnable 2D feature maps, one per canonical coordinate plane. As summarized in work on 3D-aware generation, the plane functions may be written as $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$, $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$, and $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$, with a 3D query point $p=(x,y,z)$ mapped to plane features through orthogonal projection. A canonical aggregated feature is therefore

$F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$

where bilinear interpolation over the four nearest texels is the usual sampling rule on each plane (Jia et al., 21 Jul 2025).

The representation is continuous in the sense that arbitrary 3D coordinates are queried through interpolation rather than by restricting evaluation to voxel centers. Export3D, for example, defines the tri-plane as $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$ with $C=32$ and $YZ$0 for $YZ$1 images, and maps a canonical point in $YZ$2 to each plane before bilinear lookup (Ki et al., 2024). NumGrad-Pull similarly normalizes query points to a canonical cube, samples three planes $YZ$3, $YZ$4, $YZ$5 by bilinear interpolation, and decodes the result with a lightweight MLP to obtain an SDF value (Cui et al., 2024).

Aggregation is not standardized across the literature. PET-NeuS concatenates the three sampled plane features and decodes the concatenated vector (Wang et al., 2023). NumGrad-Pull, VOODOO 3D, and MUTE-SLAM use element-wise summation across planes before decoding (Cui et al., 2024, Tran et al., 2023, Yan et al., 2024). Export3D instead averages the three plane features,

$YZ$6

and feeds the result into separate decoder heads for density and color (Ki et al., 2024). This variation is not merely cosmetic: it changes channel dimensionality, decoder burden, and the extent to which the three planes retain separate semantic roles.

2. Decoding and rendering paradigms

The most common use of tri-planes is as a volumetric feature field for radiance rendering. In EG3D-style pipelines as described by later work, a compact MLP decodes the aggregated plane feature into density and color, optionally conditioned on view direction:

$YZ$7

Rendering then follows the standard NeRF-style alpha-compositing approximation,

$YZ$8

with points sampled along a ray $YZ$9 (Jia et al., 21 Jul 2025). OTAvatar, VOODOO 3D, Export3D, and Tri$XZ$0-plane all adopt this basic division of labor: the tri-plane stores spatial features, while a shallow decoder and a differentiable renderer map those features to images under controllable viewpoint or motion (Ma et al., 2023, Tran et al., 2023, Ki et al., 2024, Song et al., 2024).

A second major use is implicit geometry. PET-NeuS replaces a pure MLP SDF with a mixture of tri-planes and MLPs, so that the signed distance at a query point is decoded from sampled plane features rather than from coordinates alone (Wang et al., 2023). NumGrad-Pull likewise models an SDF with a tri-plane encoder plus a 3-layer fully connected decoder, then trains the field with a pull-based objective in which query points are moved toward nearest points in the input cloud using numerical SDF gradients instead of analytical gradients (Cui et al., 2024). In both cases, the tri-plane serves as a localized, high-capacity feature store for geometry rather than only for appearance.

Several systems specialize the decoder while retaining the same plane-sampling front end. MUTE-SLAM uses separate tri-plane hash encoders for TSDF and color, sums the three plane features for each property, and decodes them with two MLPs (Yan et al., 2024). SHaDe uses time-indexed tri-planes for deformation, then evaluates color in canonical space with spherical harmonics coefficients and attention over learned frequency bands rather than a conventional view MLP (Alruwayqi, 22 May 2025). VRVVC stores per-frame appearance on tri-planes but decodes density from a separate 3D grid, using deferred feature accumulation so that the color MLP is evaluated once per ray rather than once per sample (Hu et al., 2024). These variants preserve the same factorized spatial interface while changing what is predicted and where the main computational budget resides.

3. Variants of the Cartesian tri-plane

The canonical Cartesian design has been extended in multiple directions to address view consistency, multi-scale detail, and geometric expressiveness. CylinderPlane replaces Cartesian projection with cylindrical coordinates, arguing that feature ambiguity in symmetric regions of standard tri-planes causes Janus or multi-face artifacts in $XZ$1 synthesis. It therefore uses cylindrical-aligned planes such as $XZ$2–$XZ$3, $XZ$4–$XZ$5, and $XZ$6–$XZ$7, together with nested cylinders at multiple radii and seam regularization at $XZ$8, so that opposite azimuths no longer share the same feature samples (Jia et al., 21 Jul 2025).

Multi-scale and hybrid variants pursue a related objective through scale decomposition or mixed coordinate systems. Tri$XZ$9-plane builds a feature pyramid of three tri-planes, $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$0, $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$1, and $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$2, with top-down lateral connections and patch-based high-resolution rendering to recover fine facial detail in monocular head avatars (Song et al., 2024). The abstract of “Implicit Gaussian Splatting with Efficient Multi-Level Tri-Plane Representation” describes a hybrid model that integrates explicit point clouds with implicit feature embeddings through a multi-level tri-plane architecture, a level-based progressive training scheme with explicit spatial regularization, and a compression pipeline for both point clouds and 2D feature grids; the cited arXiv entry, however, does not provide the full paper text beyond the abstract, so finer implementation details are unavailable in the source block (Wu et al., 2024). HyPlaneHead departs further from the standard formulation by combining planar and spherical planes, introducing near-equal-area warping for the spherical component and a single-channel unified feature map to avoid feature penetration across convolutional channels (Li et al., 20 Sep 2025).

The broader literature also contains tri-plane-like generalizations that are not simple bilinear-query fields. TP-Mamba reshapes 3D medical features into three orthogonal plane-wise sequences—height–width, depth–width, and depth–height—and processes them with Mamba blocks before fusing them by summation, thereby using a tri-plane decomposition for efficient long-range 3D adaptation of SAM rather than for neural rendering (Wang et al., 2024). “Efficient 3D Perception on Embedded Systems via Interpolation-Free Tri-Plane Lifting and Volume Fusion” removes point-wise interpolation entirely: it constructs three 2D plane features from axis-aligned averages, then reconstructs a 3D feature volume by broadcast and summation (Lee et al., 18 Sep 2025). TriVol replaces planes with three slim volumes of finite thickness, while GA-Planes generalizes tri-planes to any combination of line, plane, and volume primitives and supports convex, semiconvex, or nonconvex training depending on the decoder design (Hu et al., 2023, Sivgin et al., 2024).

4. Application domains

Human-centered generation and animation form one of the most active application areas. Export3D uses a conditional tri-plane generator with expression conditioning and appearance-free expression pre-training for one-shot portrait animation, reporting on VFHQ same-identity metrics of PSNR $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$3, SSIM $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$4, AKD $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$5, CSIM $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$6, AED $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$7, and APD $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$8 (Ki et al., 2024). VOODOO 3D lifts source and driver frames into canonical tri-planes, predicts an expression residual tri-plane, and reports real-time operation at about $f_{XY}:\mathbb{R}^2\to\mathbb{R}^C$9 FPS per view on RTX 4090 for the single-view pipeline (Tran et al., 2023). OTAvatar uses an EG3D-style tri-plane generator modulated by an identity code and a learned motion code, and reports controllable rendering at $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$0 FPS on A100 (Ma et al., 2023). TriHuman introduces a human-tailored deformable tri-plane representation in an undeformed tri-plane texture space and reports real-time rendering at at least $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$1 fps with one-frame latency on two A100 GPUs (Zhu et al., 2023). Tri$f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$2-plane adds a feature-pyramid hierarchy and reports PSNR $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$3 and LPIPS $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$4 on self-reconstruction at $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$5 resolution (Song et al., 2024).

Surface reconstruction and geometric fitting use tri-planes as compact SDF parameterizations. PET-NeuS reports a $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$6 improvement over NeuS on Nerf-synthetic in Chamfer distance, from $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$7 to $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$8, and a $f_{YZ}:\mathbb{R}^2\to\mathbb{R}^C$9 improvement on DTU, from $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$0 to $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$1, while attributing the gain to positional encoding modulation and self-attention convolution on tri-plane features (Wang et al., 2023). NumGrad-Pull uses tri-planes plus numerical gradients for pull-based surface reconstruction and reports ABC $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$2 versus IF $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$3, DIGS $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$4, SIREN $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$5, and NP $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$6, as well as a $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$7 speedup over NP and $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$8 over IF on FAMOUS (Cui et al., 2024). In perception, TP-Mamba adapts SAM to 3D CT volumes with a tri-plane Mamba module and reports an average Dice of $f_{XZ}:\mathbb{R}^2\to\mathbb{R}^C$9 on BTCV with full training data and $p=(x,y,z)$0 with only three CT training samples, surpassing the best conventional 3D network by $p=(x,y,z)$1 points in the scarce-data setting (Wang et al., 2024).

Tri-planes are also used in mapping, robotics, remote sensing, dynamics, and compression. MUTE-SLAM represents each local sub-map with tri-plane hash encodings for TSDF and color and reports competitive tracking with average Replica surface reconstruction metrics of Depth L1 $p=(x,y,z)$2 cm, Accuracy $p=(x,y,z)$3 cm, Completeness $p=(x,y,z)$4 cm, and Completeness Ratio $p=(x,y,z)$5 (Yan et al., 2024). TriDF places high-frequency color on three $p=(x,y,z)$6 planes and delegates density to a continuous field with image-based rendering features, reporting a $p=(x,y,z)$7 speed increase over NeRF-based methods and improvements over ImMPI of $p=(x,y,z)$8 in PSNR, $p=(x,y,z)$9 in SSIM, and $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$0 in LPIPS using only three training views (Kang et al., 17 Mar 2025). SHaDe uses time-indexed tri-planes for explicit deformation in dynamic 4D reconstruction and reports D-NeRF results such as PSNR $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$1, SSIM $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$2, and LPIPS $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$3 on T-Rex (Alruwayqi, 22 May 2025). VRVVC uses tri-plane implicit residuals for inter-frame volumetric video compression and reports BDBR reductions relative to TeTriRF of $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$4 on ReRF test views and $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$5 on DNA-Rendering test views (Hu et al., 2024). In robotics, Equivariant Volumetric Grasping projects a 3D feature field onto three planes and enforces $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$6 equivariance, reporting grasp success rates of $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$7 for EquiGIGA and $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$8 for EquiIGD in the packed fixed-camera setting (Song et al., 24 Jul 2025).

5. Efficiency, training, and compression

A core reason for the adoption of tri-planes is computational scaling. NumGrad-Pull states the storage of tri-planes as $F(p)=[f_{XY}(x,y); f_{YZ}(y,z); f_{XZ}(x,z)],$9 rather than the $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$0 cost of voxel grids, with each query requiring $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$1 bilinear fetches per plane and a shallow MLP pass (Cui et al., 2024). TriDF makes the same memory argument in the few-shot remote-sensing setting, emphasizing $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$2 plane storage rather than $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$3 voxels (Kang et al., 17 Mar 2025). TriVol quantifies this trade-off against dense 3D volumes: a dense $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$4 voxel baseline used $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$5 GB and $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$6 GFLOPS, whereas TriVol with $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$7 and $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$8 used $$T=(T_{XY},T_{XZ},T_{YZ})\in\mathbb{R}^{3\times C\times(H/2)\times(W/2)}$$9 GB and $C=32$0 GFLOPS while improving PSNR from $C=32$1 to $C=32$2 on ShapeNet (Hu et al., 2023). TP-Mamba extends the efficiency argument to 3D medical adaptation, contrasting near-linear plane-wise SSM scanning with the $C=32$3 cost of full 3D attention (Wang et al., 2024).

Training strategies frequently exploit the explicit spatial structure of plane features. NumGrad-Pull uses progressive plane expansion, starting at resolution $C=32$4, then $C=32$5, then $C=32$6, with the finite-difference step size tied to the current resolution (Cui et al., 2024). Tri$C=32$7-plane combines its three-scale representation with a camera-based geometry-aware sliding-window augmentation that improves robustness beyond canonical coverage (Song et al., 2024). VOODOO 3D uses iterative joint fine-tuning of the lifting and expression modules on real and synthetic data (Tran et al., 2023). This suggests that tri-plane training is often organized around curricula, coarse-to-fine schedules, or staged optimization rather than a single monolithic fit.

Compression research uses the same factorization for rate control. VRVVC divides sequences into groups of features, encodes tri-plane and density residuals, quantizes latent variables with a learnable step size, and trains one model across multiple predefined Lagrange multipliers. It reports variable-rate points such as roughly $C=32$8 KB/frame, $C=32$9 KB/frame, and $YZ$00 KB/frame on ReRF, with corresponding test-view PSNR values of $YZ$01, $YZ$02, and $YZ$03 dB (Hu et al., 2024). The abstract of IGS likewise describes a compression-oriented multi-level tri-plane representation that can deliver high-quality rendering using only a few MBs while balancing storage efficiency and fidelity (Wu et al., 2024). In these settings, tri-planes are valued not only as neural fields but also as compressible, structured feature tensors.

6. Limitations, misconceptions, and research directions

A recurrent limitation of standard Cartesian tri-planes is feature entanglement under projection. CylinderPlane states that orthogonal Cartesian projection causes symmetric regions to share samples, so that front and back facial regions can collide on the same $YZ$04 coordinates; in $YZ$05 synthesis this yields Janus or multi-face artifacts, especially under imbalanced camera distributions (Jia et al., 21 Jul 2025). HyPlaneHead makes a closely related criticism, arguing that Cartesian projection leads to feature entanglement and mirroring artifacts, while spherical tri-plane variants introduce uneven square-to-sphere mapping and seam problems if implemented with conventional $YZ$06–$YZ$07 warping (Li et al., 20 Sep 2025). The resulting controversy is not whether tri-planes are useful, but which coordinate system and which inter-plane coupling are appropriate for full-view consistency.

Geometry-oriented applications expose different weaknesses. PET-NeuS notes that direct tri-plane usage can introduce high-frequency noise in reconstructed surfaces because of projection-induced mixing, lack of frequency constraints, and overfitting to high-frequency textures (Wang et al., 2023). NumGrad-Pull argues that analytical gradients through grid-based tri-planes are unstable at cell boundaries because rounding operations make feature access piecewise defined, which motivates its use of central-difference numerical gradients (Cui et al., 2024). TriHuman identifies collisions in surface-aligned query coordinates near edges and vertices and mitigates them by reducing the sampling band and refining the template surface (Zhu et al., 2023). These papers do not reject the representation; rather, they show that interpolation, regularization, and parameterization choices remain central to stability and fidelity.

A common misconception is that tri-plane volumetric representation is exhausted by three Cartesian feature maps with bilinear interpolation and a tiny MLP. The recent literature is substantially broader. TP-Mamba uses tri-plane decomposition as a plane-wise sequence model for volumetric segmentation rather than as a radiance field (Wang et al., 2024). Interpolation-free lifting discards point-wise interpolation in favor of broadcast-and-sum reconstruction for embedded 3D perception (Lee et al., 18 Sep 2025). GA-Planes shows that tri-planes are one point in a larger design space of line-, plane-, and volume-based tensor primitives, and that some of these representations admit convex or semiconvex training for convex inverse problems (Sivgin et al., 2024). This suggests that “tri-plane” now denotes a family of factorized volumetric representations rather than a single fixed architecture.