Compositional Splatting Strategy

• Compositional Splatting Strategy is a framework where Gaussian splatting is decomposed into interpretable components, enabling precise control over rendering, representation, and optimization.
• It separates geometric occupancy from optical opacity, allowing individual adjustments such as tailored reflection, transmission, and local texture incorporation.
• The strategy further extends to scene-level decomposition by grouping objects and parts, and employing training dynamics that enhance editability and improve empirical rendering performance.

In the cited literature, the phrase compositional splatting strategy denotes a set of Gaussian-splatting formulations in which representation, rendering, and optimization are explicitly decomposed into interpretable components rather than treated as a monolithic “one opacity, one color” process. Depending on the paper, the composition may occur at the level of primitive attributes, time-varying visibility, object- or part-wise grouping, mixed geometric primitive types, or shared global/local fields, while retaining front-to-back splatting as the computational backbone (Shi et al., 18 May 2026, Wang et al., 2024, Vilesov et al., 2023, Ge et al., 2024, Liu et al., 2024, Qu et al., 15 Jul 2025, Wang et al., 24 Nov 2025).

1. Definition and conceptual scope

Standard Gaussian splatting is built on front-to-back compositing. In space-time 2D Gaussian Splatting, the rendering equation is written as

$$\mathbf{C}(r) = \sum_{i=1} c(x_i)\,\alpha(x_i)\ \prod_{j=1}^{i-1} (1 - \alpha(x_j)),$$

with samples sorted from near to far along the ray, and the same compositional weights reused for depth, normals, and masks (Wang et al., 2024). CG3D uses the same principle for projected 3D Gaussians,

$$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$

which makes explicit that visibility ordering and alpha composition are already intrinsic to Gaussian radiance fields (Vilesov et al., 2023).

Across the literature, however, compositionality is not limited to this alpha-compositing law. RT-Splatting uses the term for a factorization of each Gaussian into geometric occupancy and optical opacity, together with a hybrid surface–volume renderer for reflection and transmission (Shi et al., 18 May 2026). Space-time 2D Gaussian Splatting uses it for a compositional opacity deformation strategy that preserves binary canonical opacities while allowing per-frame visibility modulation (Wang et al., 2024). CG3D, CompGS, and DecompDreamer use it for object-wise Gaussian organization and relation-aware optimization in text-to-3D (Vilesov et al., 2023, Ge et al., 2024, Nath et al., 15 Mar 2025). GOC and GaussianBlock use it for object-centric and part-aware scene decomposition (Liu et al., 2024, Jiang et al., 2024). MP-GS uses it to splat non-elliptical primitives such as Gaussian lines and Gaussian triangles by composing vertex-based Gaussian contributions within one alpha-blending pipeline (Qu et al., 15 Jul 2025). NTS uses it to combine a coarse Gaussian scaffold with a shared global neural field that predicts local texture fields per primitive (Wang et al., 24 Nov 2025).

A concise view of these usages is given below.

Level of composition	Mechanism	Representative papers
Primitive attributes	Occupancy/opacity factorization; local RGBA fields; heterogeneous splats	(Shi et al., 18 May 2026, Wang et al., 24 Nov 2025, Daniels et al., 18 Nov 2025)
Time-varying visibility	Canonical opacity plus deformation-dependent per-frame opacity	(Wang et al., 2024)
Object or part structure	Union of disjoint Gaussian sets; semantic or part bindings	(Vilesov et al., 2023, Ge et al., 2024, Nath et al., 15 Mar 2025, Liu et al., 2024, Jiang et al., 2024, Sun et al., 16 Apr 2026)
Geometric primitive type	Ellipses, lines, and triangles in one splatting framework	(Qu et al., 15 Jul 2025)
Optimization schedule	Entity/scene alternation, relation-first refinement, gradient gating	(Shi et al., 18 May 2026, Ge et al., 2024, Nath et al., 15 Mar 2025)

A theoretical generalization is provided by "Splat Regression Models" (Daniels et al., 18 Nov 2025). There, a splat model is written as

$$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$

and for discrete support,

$$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$

For Gaussian mother splats, this recovers the functional form underlying Gaussian Splatting (Daniels et al., 18 Nov 2025). This suggests that compositional splatting can be interpreted both as alpha composition in rendering and as composition of heterogeneous anisotropic localized basis functions.

2. Primitive-level factorization and local compositionality

A major line of work makes each splat internally compositional. RT-Splatting starts from the observation that a single opacity $\alpha$ plays two roles at once: a geometric role and an optical role. Its central contribution is to disentangle each Gaussian’s geometric occupancy from its optical opacity. It introduces geometric occupancy $\sigma \in [0,1]$, optical opacity $\alpha \in [0,1]$, and their product

$\alpha_\text{eff} = \sigma \alpha$

as the effective opacity used for volumetric compositing (Shi et al., 18 May 2026). The same Gaussian set is therefore interpreted both as a surface, via $\sigma$, and as a volume, via $\alpha_\text{eff}$. Reflection is extracted from occupancy-based first-hit probabilities

$$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$0

while transmission is accumulated with effective opacity. The final color is explicitly composed as

$$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$1

with

$$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$2

In this formulation, representation-level and rendering-level factorization coincide (Shi et al., 18 May 2026).

Neural Texture Splatting introduces a different primitive-level factorization. Each Gaussian retains its center, covariance, opacity, and base color, but also receives a local RGBA texture field $$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$3. Given world-space query point $$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$4, local coordinates are

$$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$5

and the local field returns

$$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$6

Instead of storing independent texture parameters for every splat, NTS predicts per-splat local tri-planes from a shared global neural field represented as a tri-plane plus decoder (Wang et al., 24 Nov 2025). The primitive therefore becomes a composition of explicit Gaussian geometry, globally shared latent structure, and local texture refinement.

MP-GS generalizes primitive type itself. Rather than using only Gaussian ellipses, it introduces Gaussian lines and Gaussian triangles. A line or triangle is represented by its vertices plus a shared covariance, the vertices are projected as ordinary Gaussian ellipses, the line or triangle boundary is re-sketched in the image plane from projected vertices and common tangents between boundary ellipses, and the per-pixel fading parameter $$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$7 is redefined piecewise so that the primitive behaves like a composite of overlapping Gaussian contributions (Qu et al., 15 Jul 2025). For triangles, the interior may use $$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$8, vertex regions use Gaussian-like falloff, and edge quadrangles interpolate between them. The global compositing equation is unchanged; what changes is the construction of $$\mathcal{R}\left\{ \mathbf{p}, \mathbf{M}, \theta \right\} = \sum_{i \in \theta} \mathbf{c}_i \,\alpha_i \prod_{j=1}^{i-1} (1-\alpha_j),$$9 for each primitive (Qu et al., 15 Jul 2025).

In the abstract framework of Splat Regression Models, this local compositionality is generalized further. Each splat has its own center $$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$0, linear transform $$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$1, and output vector $$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$2, so heterogeneity and anisotropy are intrinsic: $$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$3 This is a compositional model in which the output is the sum of localized contributions, and Gaussian splatting appears as a special case rather than as the only permissible primitive family (Daniels et al., 18 Nov 2025).

3. Scene-level decomposition: objects, parts, and semantic components

Another major usage of compositional splatting organizes scenes as unions of explicit Gaussian subsets. CG3D represents a scene as a set of objects $$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$4, each object being its own Gaussian radiance field, together with pairwise interaction transforms $$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$5 (Vilesov et al., 2023). Object Gaussians are transformed by

$$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$6

and all transformed Gaussians are concatenated into a single compositional Gaussian radiance field (Vilesov et al., 2023). Compositionality here is geometric and explicit: separate object fields, explicit transforms, and scene-level rendering by union.

CompGS adopts an entity-structured Gaussian scene. Each segmented entity is reconstructed into its own Gaussian set

$$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$7

Its compositional strategy comes from transferring 2D compositionality to 3D initialization, keeping entity-wise Gaussian subsets, and optimizing alternately at the entity level and at the full-scene level (Ge et al., 2024). DecompDreamer uses an analogous object-wise construction,

$$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$8

but derives objects, attributes, and relationships from a VLM-generated scene graph, and trains with object prompts and edge prompts (Nath et al., 15 Mar 2025).

Object-centric reconstruction uses the same principle for real scenes rather than generative scenes. GOC augments each Gaussian with a learnable semantic feature vector $$f_\mu(x) \;=\; \mathbb{E}_{(v,\rho_{A,b})\sim \mu}\big[ v\, \rho_{A,b}(x)\big],$$9, renders features by

$$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$0

and predicts per-pixel instance probabilities with an MLP followed by a segmentation loss (Liu et al., 2024). After training, the Gaussian set is partitioned into object-wise subsets, enabling per-object point clouds, meshes, and completion (Liu et al., 2024). GaussianBlock pushes decomposition to the part level. A scene is represented as

$$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$1

where each GaussianBlock combines a superquadric primitive $$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$2 with a bound Gaussian subset $$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$3 (Jiang et al., 2024). The structure is explicitly part-aware and editable.

Compositional head avatars provide a component-specific variant. "One-shot Compositional 3D Head Avatars with Deformable Hair" separates the avatar into a bald head Gaussian set and a hair Gaussian set,

$$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$4

with different deformation models for the two subsets and joint rendering in a unified 3DGS scene (Sun et al., 16 Apr 2026). This makes the composition semantic and kinematic at the same time: face and skull are FLAME-rigged, while hair is simulated.

4. Rendering and optimization as compositional processes

In several papers, compositionality is located primarily in the training dynamics rather than only in the scene parameterization. Space-time 2D Gaussian Splatting is explicit on this point. It treats compositional splatting as standard front-to-back alpha composition of 2D Gaussian surfels, but extends it into space-time with canonical surfels, a 4D geometry deformation field, and an opacity deformation field

$$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$5

The location of the product and nonlinearity is the key difference from additive or multiplicative opacity updates. The paper describes this as maintaining a stable, clear binary distribution of canonical opacities while allowing transient visibility modulation over time (Wang et al., 2024).

RT-Splatting introduces a training-time composition problem: which branch should explain residuals at a given pixel when reflection and transmission are jointly optimized. Its answer is Specular-Aware Gradient Gating. The local gate is

$$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$6

with $$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$7, and the transmission-branch gradient is scaled by $$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$8 (Shi et al., 18 May 2026). The forward pass is unchanged, but the backward pass becomes branch-specific. This is a compositional mechanism at the optimization level: the global image loss is spatially repartitioned into different effective supervision strengths for reflection and transmission (Shi et al., 18 May 2026).

Text-to-3D systems use comparable schedule-level decomposition. CompGS alternates entity-level SDS and composition-level SDS, so the same scene is optimized sometimes as isolated Gaussian subsets and sometimes as a full union. It also introduces Volume-Adaptive Optimization, temporarily rescaling each entity into a standardized volume before its entity-level step (Ge et al., 2024). DecompDreamer uses a progressive schedule: an early relationship-first regime dominated by edge-wise joint optimization, and a later object refinement regime dominated by targeted object losses and object-view-aware prompts (Nath et al., 15 Mar 2025). In both cases, compositionality is realized by deciding which Gaussian subsets receive gradients under which textual conditions.

Semantic reconstruction papers employ analogous multi-term objectives. GOC jointly optimizes color, monocular depth priors, depth-induced normal consistency, multi-view photometric reprojection, multi-view geometry consistency, and segmentation,

$$f_\mu(x)=\sum_{i=1}^k w_i\, v_i \,\rho\big(A_i^{-1}(x-b_i)\big)\,|\det A_i^{-1}|.$$9

with $\alpha$0 (Liu et al., 2024). GaussianBlock similarly separates primitive-level semantic organization from Gaussian-level photometric refinement: Stage 1 optimizes superquadrics with reconstruction and Attention-guided Centering losses, while Stage 2 refines bound Gaussians with RGB and positional regularization (Jiang et al., 2024). These formulations suggest that a compositional splatting strategy often includes an explicit decomposition of supervision, not merely of geometry.

5. Applications, editability, and empirical behavior

The most direct application is reflection–transmission rendering. RT-Splatting targets thin semi-transparent specular surfaces such as glass and plastic films, where standard 3DGS tends to blur reflections or overly occlude transmission. Its factorized occupancy/opacity representation, hybrid renderer, and gradient gating yield state-of-the-art performance on challenging semi-transparent scenes and naturally enable component-wise editing: changing roughness $\alpha$1, transmissivity ratio $\alpha$2, optical opacity $\alpha$3, reflection strength, or scattered color without altering the whole scene model (Shi et al., 18 May 2026).

A second application is dynamic surface reconstruction under occlusion. Space-time 2D Gaussian Splatting reports that its compositional opacity deformation strategy reduces the surface recovery of occluded areas. On CMU Panoptic, the ablation on opacity strategy gives average Chamfer values of 15.2 mm without opacity deformation, 15.3 mm for multiplicative opacity deformation, 15.2 mm for additive deformation, and 14.6 mm for the compositional formulation, with a similarly compact point count of about $\alpha$4 (Wang et al., 2024). The same work reports average Chamfer 14.6 mm versus 15.7 mm for SDFFlow and 4DGS in Table 2 (Wang et al., 2024).

Generative compositionality is a third application. CompGS evaluates on T$\alpha$5Bench multiple-objects and reports Quality 54.2, Alignment 37.9, and Average 46.1, outperforming VP3D by +5.1 in Quality and +6.4 in Alignment (Ge et al., 2024). Its ablation reports an average score drop from 45.9 to 20.8 without 2D compositional Gaussian initialization, to 41.0 without Decomposed Optimization, and to 43.6 without Volume-Adaptive Optimization (Ge et al., 2024). DecompDreamer reports on 30 compositional prompts CLIP score 34.5 versus next best 32.34, Pick-A-Pic 22.5% versus next best about 17%, and user study preference 54% versus next best 18.3% (Nath et al., 15 Mar 2025). CG3D, although earlier, established the same object-wise Gaussian field plus interaction-transform pattern for physics-aware multi-object composition (Vilesov et al., 2023).

Object-centric reconstruction and editing form another cluster. GOC reports approximately 1.1 hours per synthetic scene versus 18–21 hours for ObjectSDF++ and RICO, and supports object-level semantic mesh visualization, rearrangement, and editing (Liu et al., 2024). GaussianBlock emphasizes direct manipulation of semantic parts—translation, rotation, scaling, deletion, duplication, and recombination of GaussianBlocks—while maintaining high-quality rendering by binding Gaussians to superquadric primitives (Jiang et al., 2024). The head-avatar formulation extends this principle to animation: a complete avatar is formed from a FLAME-rigged bald component and a cage-simulated hair component, producing dynamic animations under head motion, gravity effects, and expressions (Sun et al., 16 Apr 2026).

A final application is surface reconstruction with mixed primitives. MP-GS reports DTU Chamfer distance 0.46 versus 0.58 for GeoFieldSplat, 0.80 for 2D-GS, and 0.74 for GOF, and Tanks & Temples F1 0.52 versus 0.30 for 2D-GS and 0.46 for GOF (Qu et al., 15 Jul 2025). Its ablations give Chamfer 0.61 without Gaussian lines, 0.64 without Gaussian triangles, 0.56 without the proposed initialization, and 0.53 without vertex pruning (Qu et al., 15 Jul 2025). These results tie the empirical gains directly to the compositional primitive formulation.

6. Distinctions, limitations, and theoretical outlook

The literature suggests that compositional splatting strategy is not a single standardized operator. In some works it still means literal front-to-back alpha composition, but with additional structured variables. This is explicit in space-time 2D Gaussian Splatting, where compositional splatting is “literally the standard front-to-back alpha blending of 2D Gaussian splats,” extended by time-dependent geometry and opacity deformation (Wang et al., 2024). In RT-Splatting, by contrast, compositionality is the decomposition of one Gaussian into geometry, optics, reflection, transmission, and branch-specific gradient flows (Shi et al., 18 May 2026). In MP-GS it is the algebraic reconstruction of line and triangle behavior from vertex ellipses and tangents (Qu et al., 15 Jul 2025). In GOC, GaussianBlock, CompGS, DecompDreamer, CG3D, and compositional head avatars, it is primarily a question of how Gaussian subsets are grouped, transformed, and optimized (Liu et al., 2024, Jiang et al., 2024, Ge et al., 2024, Nath et al., 15 Mar 2025, Vilesov et al., 2023, Sun et al., 16 Apr 2026).

The limitations are correspondingly heterogeneous. RT-Splatting is designed for thin semi-transparent surfaces with negligible refraction and does not model thick glass, liquids, or strong internal caustics (Shi et al., 18 May 2026). Space-time 2D Gaussian Splatting is optimized for mostly solid, non-transparent scenes and notes that binary canonical opacities are less suited to semi-transparent materials or volumetric effects (Wang et al., 2024). CompGS reports difficulty with backgrounds and abstract entities such as grassland, sky, and water because text-guided segmentation does not provide reliable 2D priors (Ge et al., 2024). DecompDreamer depends on VLM decomposition and is still subject to SDS/FMD artifacts and limited hierarchical decomposition within a single object (Nath et al., 15 Mar 2025). GOC assumes reliable instance masks during training, and its object completion model relies only on geometry, so highly ambiguous shapes can be completed incorrectly (Liu et al., 2024). GaussianBlock is object-centric and handles backgrounds only in a simplified manner (Jiang et al., 2024).

A theoretical unification is offered by Splat Regression Models. That work separates three notions that are often conflated in practice: the model $\alpha$6, the inverse problem defined by a task-specific functional $\alpha$7, and the optimization algorithm given by Wasserstein–Fisher–Rao gradient flows on the space of mixing measures (Daniels et al., 18 Nov 2025). The resulting gradients with respect to output vectors $\alpha$8, transforms $\alpha$9, and centers $\sigma \in [0,1]$0 make explicit how splats can be created, moved, reshaped, and removed within a principled measure-theoretic geometry (Daniels et al., 18 Nov 2025). This suggests a broad interpretation of compositional splatting: not merely a rendering heuristic, but a design pattern in which localized primitives are endowed with structured internal factors, organized into semantic or physical subsets, and optimized through explicitly decomposed objectives while remaining compatible with efficient splat-based rendering.