Triangle-Aware Gaussian Manipulation

• Triangle-aware Gaussian manipulation is a design space that constrains Gaussian primitives using explicit triangle structures for sharp boundary rendering and deformation control.
• It encompasses methods such as BG-Triangle, Mani-GS, and PLANING, each leveraging triangles differently to enhance rendering fidelity, editing, or streaming reconstruction.
• By coupling triangle geometry with Gaussian rendering, these approaches achieve improved discontinuity preservation, efficient level-of-detail control, and robustness under deformation.

Triangle-aware Gaussian manipulation denotes a family of methods in which Gaussian primitives are generated from, bound to, or anchored by triangle-based structure rather than treated as unconstrained splats. In current 3D graphics and reconstruction research, the central motivation is to combine the differentiability and rendering efficiency of Gaussian primitives with the explicit boundary structure, controllable deformation, or geometric regularity supplied by triangles. Representative formulations include Bézier Gaussian Triangle (BG-Triangle), which generates Gaussian sub-primitives from Bézier triangles for differentiable rendering; Mani-GS, which binds 3D Gaussian Splatting to a triangular mesh for controllable manipulation; and PLANING, which loosely couples explicit triangle primitives with neural Gaussians for streaming 3D reconstruction (Wu et al., 18 Mar 2025, Gao et al., 2024, Jiang et al., 29 Jan 2026).

1. Conceptual basis

A common starting point for triangle-aware Gaussian methods is the observation that differentiable rendering systems often rely on smooth, probabilistic proxies such as volumes or Gaussian primitives, and therefore struggle to preserve sharp edges because they lack explicit boundary definitions. Triangle-aware formulations address this by introducing an explicit triangle domain or mesh scaffold while retaining Gaussian-based rendering or appearance modeling (Wu et al., 18 Mar 2025).

Across the recent literature, triangles serve three distinct but related roles. In BG-Triangle, they are vectorized, explicit, and resolution-independent geometric primitives from which Gaussian sub-primitives are generated on the fly for rendering. In Mani-GS, a triangular mesh is a manipulable proxy that defines local coordinate frames for Gaussians and drives their self-adaptation under deformation. In PLANING, triangles are the primary geometric scaffold, while neural Gaussians encode appearance in a decoupled manner (Wu et al., 18 Mar 2025, Gao et al., 2024, Jiang et al., 29 Jan 2026).

This division of labor is significant because it reframes Gaussian representations as triangle-conditioned rather than purely point-based. A plausible implication is that triangle awareness is not a single algorithmic trick but a representational principle: explicit triangle structure is introduced wherever unconstrained Gaussian overlap would otherwise blur boundaries, destabilize deformation, or create structural redundancy.

2. Representational forms

The recent literature contains three principal representational patterns for triangle-aware Gaussian systems.

Framework	Triangle role	Gaussian role
BG-Triangle	Bézier triangle primitive	Per-pixel Gaussian sub-primitive
Mani-GS	Manipulable triangular mesh	Bound 3DGS primitive with self-adaptation
PLANING	Explicit geometry scaffold	Neural appearance primitive

BG-Triangle combines Bézier triangles with Gaussian-based probabilistic models. Mani-GS extracts or supplies a triangular mesh and initializes Gaussians in triangle-local coordinates. PLANING anchors each Gaussian to a parent triangle while optimizing triangle geometry and Gaussian appearance separately (Wu et al., 18 Mar 2025, Gao et al., 2024, Jiang et al., 29 Jan 2026).

These variants differ primarily in how tightly the Gaussian is constrained by triangle structure. BG-Triangle keeps Gaussians within their originating triangle and computes them per pixel. Mani-GS keeps learned local attributes invariant in triangle space but recomputes global position, rotation, and scale after deformation. PLANING uses a loosely coupled representation in which geometry and appearance are modeled in a decoupled manner, although Gaussians still retain explicit triangle anchors (Wu et al., 18 Mar 2025, Gao et al., 2024, Jiang et al., 29 Jan 2026).

3. BG-Triangle and differentiable rendering

BG-Triangle formalizes a hybrid representation in which a degree-$n$ Bézier triangle surface is defined over barycentric coordinates $(u,v,w)$, with $u+v+w=1$, by

$$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$

where

$$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$

Each primitive encodes geometry as well as per-corner attributes such as color or SH coefficients. For rendering, each foreground pixel obtains a Gaussian sub-primitive determined from the primitive index and barycentric coordinates, with attributes interpolated from triangle attributes. The key constraint is that these Gaussians are always within their originating triangle (Wu et al., 18 Mar 2025).

Its central rendering mechanism is a discontinuity-aware alpha blending scheme intended to reduce uncertainties at object boundaries. After rasterization, triangle pixel boundaries are detected via neighboring pixel label difference, boundary pixels are stored, and a boundary-aware blending weight $w(\mathbf{q})$ is computed with a smooth transition function

$$\gamma(d;\sigma)=\min\left(2^{\frac{d}{\sigma}-1},1\right).$$

Each Gaussian’s pixel alpha is then

$$\alpha(\mathbf{q})=o\cdot w(\mathbf{q})\cdot e^{-\frac{1}{2}(\mathbf{q}-\mu)^\top \Sigma^{-1}(\mathbf{q}-\mu)},$$

and alpha blending proceeds as in standard point-based rendering:

$$C=\sum_{i=0}^{n-1}T_i\alpha_i c_i + T_n c_{\text{bg}},\qquad T_i=\prod_{j=0}^{i-1}(1-\alpha_j),\quad T_0=1.$$

The paper states that all steps, including barycentric mapping, $w(\mathbf{q})$, and $(u,v,w)$0, are smooth and differentiable, so backpropagation can adjust triangle control points to align with scene edges (Wu et al., 18 Mar 2025).

BG-Triangle also introduces adaptive splitting and pruning for level-of-detail control. Densification is triggered by high gradient amplitude on a triangle’s control points or by an edge prior derived from projected image edge gradients mapped to a triangle; splitting subdivides a triangle into four by inserting edge midpoints and re-assigning attributes. Pruning uses visibility, visible area, area, and aspect ratio criteria. The reported outcome is comparable rendering quality to 3DGS with superior boundary preservation and a much smaller number of primitives than alternatives, which the paper attributes to vectorized graphics primitives and adaptive refinement (Wu et al., 18 Mar 2025).

4. Mesh-guided manipulation in Mani-GS

Mani-GS addresses a different problem: manipulating Gaussian splats while maintaining rendering quality. Its approach is to use a triangular mesh to manipulate 3DGS directly with self-adaptation. A mesh is extracted from 3DGS using methods such as Screened Poisson Reconstruction or Marching Cubes, and each triangle becomes the local coordinate frame for Gaussian initialization (Gao et al., 2024).

For a triangle $(u,v,w)$1 with vertices $(u,v,w)$2, Gaussians are placed at predefined barycentric coordinates $(u,v,w)$3, where $(u,v,w)$4, with initial mean

$(u,v,w)$5

The local triangle axes are defined from an edge direction, the triangle normal, and their cross product, producing a rotation matrix $(u,v,w)$6. Local Gaussian attributes $(u,v,w)$7 are learned in this triangle-local space (Gao et al., 2024).

The defining self-adaptation mechanism keeps local attributes invariant in triangle space during manipulation while recomputing global Gaussian properties from the updated mesh:

$(u,v,w)$8

Here $(u,v,w)$9 is a shape adaptation vector derived from triangle edge lengths. The paper presents this as the basis for large deformations, local manipulations, and soft body simulations, and emphasizes that Gaussians are not constrained exactly to triangle surfaces but are allowed to be offset and flexible in triangle-local space, which increases tolerance to inaccurate meshes (Gao et al., 2024).

The reported quantitative results on the NeRF Synthetic Dataset are PSNR $u+v+w=1$0, SSIM $u+v+w=1$1, and LPIPS $u+v+w=1$2 for Mani-GS, compared with PSNR $u+v+w=1$3, SSIM $u+v+w=1$4, and LPIPS $u+v+w=1$5 for SuGaR. The paper also reports rendering up to 300 FPS with $u+v+w=1$6K points. Ablations state that fixed attachment fails under deformation if the mesh is not perfectly accurate, whereas fixed offset helps statically but causes misalignment after deformation; the triangle-aware, self-adapting strategy is presented as the alternative that preserves visual and numerical fidelity (Gao et al., 2024).

5. Loosely coupled streaming reconstruction in PLANING

PLANING moves triangle-aware Gaussian manipulation into online reconstruction. It introduces a hybrid representation that loosely couples explicit triangle primitives with neural Gaussians, with triangles modeling geometry and neural Gaussians modeling appearance. The paper states that this decoupling supports an online initialization and optimization strategy that separates geometry and appearance updates (Jiang et al., 29 Jan 2026).

A triangle primitive is defined by its three vertices $u+v+w=1$7, with a local frame centered at the barycenter $u+v+w=1$8, tangent vectors, and unit normal. Each Gaussian $u+v+w=1$9 is associated with a parent triangle $$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$0, and its absolute position is expressed as

$$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$1

Scale and rotation are decoded from fused triangle and Gaussian features:

$$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$2

The paper further specifies separate parameter sets and losses for geometry and appearance, optional appearance-to-geometry gradient flow, and asynchronous global-map synchronization when poses are updated (Jiang et al., 29 Jan 2026).

For geometry, PLANING supervises triangles with multi-view depth and normal priors and an opacity entropy loss. For appearance, it uses an RGB loss combining $$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$3, SSIM, and a scale regularizer. Primitive initialization is filtered photometrically and spatially to reduce redundancy, and the system uses subdivision-aware forward rendering plus a custom CUDA rasterizer (Jiang et al., 29 Jan 2026).

The abstract reports that PLANING improves dense mesh Chamfer-L2 by $$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$4 over PGSR, surpasses ARTDECO by $$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$5 dB PSNR, and reconstructs ScanNetV2 scenes in under 100 seconds, over $$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$6 faster than 2D Gaussian Splatting, while matching the quality of offline per-scene optimization. The detailed ScanNetV2 results give Chamfer-L2 $$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$7, F-score $$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$8, and $$\mathbf{S}(u,v,w)=\sum_{i=0}^{n}\sum_{j=0}^{n-i} B^n_{i,j,k}(u,v,w)\,\mathbf{p}_{i,j,k},$$9primitives $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$0k for PLANING, compared with Chamfer-L2 $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$1, F-score $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$2, and $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$3primitives $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$4k for ARTDECO, and Chamfer-L2 $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$5, F-score $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$6, and $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$7primitives $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$8k for 2DGS. For rendering on ScanNetV2, the paper reports PSNR $$B^n_{i,j,k}(u,v,w)=\frac{n!}{i!j!k!}u^i v^j w^k,\qquad k=n-i-j.$$9, SSIM $w(\mathbf{q})$0, and LPIPS $w(\mathbf{q})$1 for PLANING, versus PSNR $w(\mathbf{q})$2, SSIM $w(\mathbf{q})$3 for ARTDECO and PSNR $w(\mathbf{q})$4, SSIM $w(\mathbf{q})$5 for 2DGS (Jiang et al., 29 Jan 2026).

6. Terminological scope and older Gaussian-triangle literature

The phrase “Gaussian triangle” has an older and distinct meaning in geometric probability and shape theory. In that literature, the object of study is a random triangle induced by Gaussian sampling, not a graphics primitive used for rendering. Finch studies pinned, staked, and anchored Gaussian triangles, deriving explicit joint densities for side lengths and angles, acuteness probabilities, and Rice-distributed marginals in several settings (Finch, 2014).

A particularly notable result is that under the side constraint $w(\mathbf{q})$6, the bivariate angle density coincides with that for Gaussian triangles in three-dimensional space. Finch gives the common angle density as

$w(\mathbf{q})$7

for $w(\mathbf{q})$8, $w(\mathbf{q})$9, and $$\gamma(d;\sigma)=\min\left(2^{\frac{d}{\sigma}-1},1\right).$$0 (Finch, 2014).

Edelman and Strang place random triangles from six independent standard Gaussian coordinates into shape theory, showing that after translation reduction and reformulation as a $$\gamma(d;\sigma)=\min\left(2^{\frac{d}{\sigma}-1},1\right).$$1 matrix problem, the induced triangle shapes are uniformly distributed on a hemisphere. They also report that the probability a random triangle is acute is $$\gamma(d;\sigma)=\min\left(2^{\frac{d}{\sigma}-1},1\right).$$2, and hence the probability it is obtuse is $$\gamma(d;\sigma)=\min\left(2^{\frac{d}{\sigma}-1},1\right).$$3 (Edelman et al., 2015).

This older literature matters for terminology. A common misconception is to read “triangle-aware Gaussian” as referring exclusively to Gaussian splatting pipelines. In fact, the term joins two historically different traditions: one studies Gaussian-induced distributions on triangle shape, and the other uses triangle structure to constrain Gaussian rendering or deformation.

7. Recurrent design patterns, distinctions, and limitations

Recent triangle-aware Gaussian systems share a common design intent but differ sharply in how the triangle constrains the Gaussian. BG-Triangle states that Gaussians are always within their originating triangle and that their attributes are smoothly interpolated from triangle attributes. Mani-GS states that Gaussians are not constrained exactly to triangle surfaces and may remain offset in triangle-local space. PLANING states that appearance gradients from Gaussians can optionally update triangle geometry for refinement, but the two are never strictly locked as in surfel-based methods (Wu et al., 18 Mar 2025, Gao et al., 2024, Jiang et al., 29 Jan 2026).

These differences correspond to different operational targets. BG-Triangle is optimized for resolution-independent differentiable rendering with boundary preservation and adaptive level of detail. Mani-GS is optimized for controllable editing, including large deformations, local manipulations, and soft body simulations. PLANING is optimized for streaming reconstruction, structural clarity, and reduced redundancy under online updates (Wu et al., 18 Mar 2025, Gao et al., 2024, Jiang et al., 29 Jan 2026).

The limitations described in the papers are likewise task-specific. Mani-GS notes that extreme non-rigid deformations of highly inaccurate meshes can still cause distortions. BG-Triangle relies on explicit boundary detection and threshold-based densification and pruning. PLANING depends on asynchronous pose correction, subdivision-aware rendering, and custom rasterization to maintain consistency and efficiency (Gao et al., 2024, Wu et al., 18 Mar 2025, Jiang et al., 29 Jan 2026).

Taken together, these works suggest a continuum rather than a single doctrine. At one end are strictly triangle-domain Gaussians used to preserve discontinuities; at the other are loosely anchored Gaussians that inherit only enough triangle structure to remain editable or geometrically organized. This suggests that “triangle-aware Gaussian manipulation” is best understood as a design space for regulating Gaussian freedom through triangle structure, with the degree of coupling chosen according to whether the dominant requirement is sharp rendering, deformation control, or scalable reconstruction.